Hi Ryan and Sébastien,
I'm sorry for late.
When I applied the source(
https://gist.github.com/slorber/7199403), it
worked well(Thanks to Sébastien!).
But I think that we can devise a better logic(I think that we can reuse
MultipartBody class which has multipart protocol, so we don't need to
consider the content type and boundary for multipart and can reduce buffer
copies in FeedBodyGeneratorOutputStream).
I think current multipart's logic of AHC with grizzly provider has two
problems.
1. Before writing, it allocates all source's buffers unfortunately.( I
already met OOM before actual writing when I tested).
In GrizzlyAsyncHttpProvider's PartsBodyHandler.
---
mre.writeRequest(o); // OOM before actual writing!
…
ctx.write(content, …);
---
It can be improved with using MultipartBody(not using
MultipartRequestEntiry#writeRequest()).
Here is a simple idea.
In GrizzlyAsyncHttpProvider's PartsBodyHandler.
---
final Body bodyLocal = new MultipartBody(request.getParts(), contentType,
String.valueOf(contentLength));
…
while(!last) {
Buffer buffer = mm.allocate(MAX_CHUNK_SIZE);
buffer.allowBufferDispose(true);
final long readBytes = bodyLocal.read(buffer.toByteBuffer());
if (readBytes > 0) {
...
} else {
buffer.dispose();
if (readBytes < 0) {
last = true;
buffer = Buffers.EMPTY_BUFFER;
} else {
...
}
}
…
ctx.write(content, …);
}
return true;
---
2. When I modified above 1 with MultipartBody like netty provider, I also
met OOM in writing because the write buffer queue will be able to be full.
I think this case is reasonable. For this problem, write-flow-control is
needed like BaseFeeder, SimpleFeeder and MultipartBodyGeneratorFeeder.
I think that it is better that the user can use the AHC without considering
grizzly provider specific setting for multipart.
ex) Maybe PartsBodyHandler can be improved with MultipartBody and BaseFeeder
logic.
In addition, multipart can have the large file so it will be better if
grizzly provider can also support the FileTransfer in PartsBodyHandler like
FileBodyHandler of GrizzlyAsyncHttpProvider.java.
What do you think? :-)
Regards,
Bongjae Chang
From: Ryan Lubke <ryan.lubke_at_oracle.com>
Reply-To: <users_at_grizzly.java.net>
Date: Tuesday, October 29, 2013 9:23 AM
To: <users_at_grizzly.java.net>
Subject: Re: Upload a large file without oom with Grizzly
I'll take a look at it later this week and see what's possible.
Sébastien Lorber wrote:
>
> Hi,
>
> Bongjae, this is code on my own project and is not part of AHC.
>
> You will find it here:
> https://gist.github.com/slorber/7199403
> It requires the last AHC release.
>
> Ryan if you want you can try to integrate it in AHC.
> Maybe it could even be made the default behavior for Grizzly based AHC
> multipart requests, and will reduce the memory footprint to a lot of users in
> the next AHC release.
>
>
> It still work fine for us until now, we had the first load tests done by our
> client.
> We don't know the test scenarios and many things were changed in addition to
> this part of the project, but at least we have good results until now and
> we'll increase the load soon.
> Won't be able to provide another feedback as i'm changing job in 3 days :)
>
>
> Bye
>
>
>
>
>
>
>
> 2013/10/28 Bongjae Chang <bongjae.chang_at_gmail.com>
>> Hi Ryan and Sébastien,>>
>> I met the same issue(OOM) while I sent the large file(about 1G size) with
>> multipart in AHC.
>> When I reviewed AHC's sources, I couldn't find MultipartBodyGeneratorFeeder
>> class yet.
>>
>> Is there any updates? or where can I get the stable reference?
>>
>> Thank in advance!
>>
>> Regards,
>> Bongjae Chang
>>
>>
>> From: Sébastien Lorber <rber.sebastien_at_gmail.com>
>> Reply-To: <users_at_grizzly.java.net>
>> Date: Wednesday, September 25, 2013 5:47 PM
>> To: "users_at_grizzly.java.net" <users_at_grizzly.java.net>
>> Subject: : Upload a large file without oom with Grizzly
>>
>> Thanks Ryan, will use that release :)
>>
>> No problem until now, will tell you what happens during our load tests
>>
>>
>> 2013/9/24 Ryan Lubke <ryan.lubke_at_oracle.com>
>>> Sébastien,
>>>
>>> AHC 1.7.20 has been released. It's up on the oss.sonatype.org
>>> <http://ss.sonatype.org> maven repository and will be sync'd with central
>>> soon-ish.
>>>
>>> -rl
>>>
>>> Ryan Lubke wrote:
>>>>
>>>>
>>>> Sébastien Lorber wrote:
>>>>> Yes it works nicely :)
>>>>>
>>>>> Thanks Ryan I think everything is now working for me
>>>>>
>>>> You're welcome. Let us know if you run into anything else, or if you have
>>>> idas on improvements, etc.
>>>>
>>>>
>>>>>
>>>>
>>>>> 2013/9/20 Ryan Lubke <ryan.lubke_at_oracle.com>
>>>>>> Just followin up for closure here. Pool issue has been confirmed as
>>>>>> resolved.
>>>>>> >>>>>> Sébastien Lorber wrote:
>>>>>>> Ok thanks :)
>>>>>>>
>>>>>>> Take your time
>>>>>> We'll keep using no connection pool (as we used to do until now anyway,
>>>>>>> and we could probably continue with this setting for a while because it
>>>>>>> passed our load test goals)
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>> 2013/9/19 Ryan Lubke <ryn.lubke_at_oracle.com>
>>>>>>> Sorry, been under the weather. I'll be investigaing this issue this
>>>>>>> morning.
>>>>>>>
>>>>>>> Stay tuned.
>>>>>>>
>>>>>>> Sébastien Lorber wrot:
>>>>>>> So what I've found is that:
>>>>>>>
>>>>>>>
>>>>>>> private fnal Connection.CloseListener listener = new
>>>>>>> Connection.CloseListener() {
>>>>>>> @Override
>>>>>>> public void onClosed(Connection connection,
>>>>>>> Connection.CloseType closeType) throws IOException {
>>>>>>> if (closeType == Connection.CloseType.REMOTEL) {
>>>>>>> if (LOG.isInfoEnabled()) {
>>>>>>> LOG.info("Remote closed connecion ({}).
>>>>>>> Removing from cache", connection.toString());
>>>>>>> }
>>>>>>> }
>>>>>>> GrizzlyConnectionsPool.this.removeAll(connection);
>>>>>>> }
>>>>>>> };
>>>>>>>
>>>>>>>
>>>>>>> public boolean removeAll(Connection connection) {
>>>>>>> >>>>>>> if (connection == null || closed.get()) {
>>>>>>> return false;
>>>>>>> }
>>>>>>> connection.removeCloseListener(listener);
>>>>>>> bo isRemoved = false;
>>>>>>> for (Map.Entry<String, DelayedExecutor.IdleConnectionQueue>
>>>>>>> entry : connectionsPool.entrySet()) {
>>>>>>> booleanremoved = entry.getValue().remove(connection);
>>>>>>> isRemoved | removed;
>>>>>>> }
>>>>>>> return isRemoved;
>>>>>>>
>>>>>>> }
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>> Whe the connection is closed remotely, the connection that we try to
>>>>>> remove is never in the cache, thus isRemoved = false
>>>>>>>
>>>>>>> I guess it has something to do with the "connection dicrimination" you
>>>>>>> previously fixed and I didn't really understand :)
>>>>>>>
>>>>>>> Do you have an idea on this problem?
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>> By the way, I just tested again WITHOUT the FeedableBodyGenerator.
>>>>>>> I run batches of 10 concurrent uploads ith a max 20 connection pool.
>>>>>>> I run these 10 concurrent uploads multiple times.
>>>>>>> It works fine, but if just after the request I add a Thread.sleep(15000)
>>>>>>> then it seems to lead to the same problem:
>>>>>>>
>>>>>>> @Test
>>>>>>> public void do_test_without_ahc_streaming() throws xception {
>>>>>>> List<Response> responses = runConcurrentl(CONCURRENT_UPLOADS,new
>>>>>>> Callable<Response>() {
>>>>>>> @Overrde
>>>>>>> public Response call() throws Exception {
>>>>>>> reurn uploadSingleDocumentWithoutStreaming();
>>>>>>> }
>>>>>>> });
>>>>>> runConcurrently(CONCURRENT_UPLOADS,new Callable<Response>() {
>>>>>>> @Override
>>>>>>> public Response call() throws Exception {
>>>>>> return uploadSingleDocumentWithoutStreaming();
>>>>>>> }
>>>>>>> });
>>>>>>> Thread.sleep(15000);
>>>>>>> runConcurrently(CONCURRENT_UPLOADS,new Callable<Respons>() {
>>>>>>> @Override
>>>>>>> public Response call() throws Exception {
>>>>>>> return uploadSingleDocumentWithoutStreaming();
>>>>>>> }
>>>>>>> });
>>>>>>> Thread.sleep(15000);
>>>>>>> runConcurrently(CONCURRENT_UPLOADS,new Callable<Response>() {
>>>>>>> @Override
>>>>>>> public Response call() throws Exception {
>>>>>>> return uploadSingleDocumentWithoutStreaming();
>>>>>>> }
>>>>>>> });
>>>>>>> Thread.sleep(15000);
>>>>>>> runConcurrently(CONCURREN_UPLOADS,new Callable<Response>() {
>>>>>>> @Override
>>>>>>> public Response call()throws Exception {
>>>>>>> return uploadSingleDocumentWithoutStreaming();
>>>>>>> }
>>>>>>> });
>>>>>>> Thread.sleep(1500);
>>>>>>> .....................................;
>>>>>>> }
>>>>>>>
>>>>>>> This permits to give some time fr the remote host to close the
>>>>>>> connection
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>> So in the end it eems that we have troubles in any case when the remote
>>>>>>> host closes the conection
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>> Btw I just saw this issue:
>>>>>>> https://github.com/AsyncHttpClient/async-http-client/pull/311
>>>>>>> I checked the code that used AHC 1.7.7 an what I can see is that the
>>>>>>> connection pool is bypassed even with the property=true
>>>>>>> I havin my logs:
>>>>>>> [main] DEBUG
>>>>>>> com.ning.http.client.providers.grzzly.GrizzlyConnectionsPool:162 -
>>>>>>> [poll] No existing queue for uri [https://digipost.orsid.com:443].
>>>>>>>
>>>>>>>
>>>>>>> So, by updating to 1.7.20-SNAPSHOT, the connection pool has been
>>>>>>> magically enabled, and itdoesn't seem to work well for me.
>>>>>>>
>>>>>>>
>>>>>>> I'll keep that poo disabled because it's the behavior of our previous
>>>>>>> applicative version that is neato production and works pretty fine,
>>>>>>> but will open a Github issue because i'm quiting my job next month and
>>>>>>> they want to track when they'll be able to use a connection pool :)
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>> 2013/9/19 Sébastien Lorber <lorber.sebastien_at_gmail.com>
>>>>>>> So it seems I can reproduce this in local.
>>>>>>>
>>>>>>> Disabling the ssl connection pooling solve the problem.
>>>>>>>
>>>>>>>
>>>>>>> We didn't have this problems with v.7.7 we previously used.
>>>>>>>
>>>>>>> I just tested with the last commits of 1.7.20-SNAPSHOT, the problem is
>>>>>>> still here.
>>>>>>> The problem doesn't appear on normal AHC requests, neither on multipart
>>>>>>> file upload when I do not use the FeedableBodyGenerator so it seems this
>>>>>>> is the new behavior of FeedableBodyGenerator that has a problem with the
>>>>>>> ssl connection pooling. Will try to investigate this.
>>>>>>>
>>>>>>> By the way, your last commits about selector/worker threads seems ok
>>>>>>> (didn't notice any problem with the ssl pool disabled)
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>> 2013/9/18 Sébastien Lorber <lorber.sebastien_at_gmail.com>
>>>>>>> Hi
>>>>>>>
>>>>>>> Unfortunatly it seems there's another problem.
>>>>>> On our test environment I often get in the logs:
>>>>>>>
>>>>>>> Caught internal server errorjava.io.IOException: Maximum pooled
>>>>>>> connections exceeded
>>>>>>> at
>>>>>>> com.ning.htp.client.providers.grizzly.GrizzlyAsyncHttpProvider$AsyncHtt
>>>>>>> pClientEventFiltr.cleanup(GrizzlyAsyncHttpProvider.java:1415)
>>>>>>> ~[async-http-client-1.7.20-2846817.jar:na]
>>>>>>> at
>>>>>>> com.ning.http.client.providers.grizzly.GrizzlyAsyncHttpProvider$AsyncHtt
>>>>>>> pClientEventFilter.onHttpPacketParsed(GrizzlyAsyncHttpProvider.java:136
>>>>>>> ) ~[async-http-client-1.7.20-2846817.jar:na]
>>>>>>> at
>>>>>>> org.glassfish.grizzly.http.HttpCodecFilter.decodeWithTransferncoding(Ht
>>>>>>> tpCodecFilter.java:1176) ~[grizzly-http-2.3.5.jar:2.3.5]
>>>>>> at
>>>>>>> org.glassfish.grizzly.http.HttpCodecFilter.handleRead(HttpCodecFiter.ja
>>>>>>> va:516) ~[grizzly-http-2.3.5.jar:23.5]
>>>>>>> at
>>>>>>> org.glassfish.grizzly.http.HttpClientFilter.handleRed(HttpClientFilter.
>>>>>>> java:161) ~[grizzly-http-2.3.5.jar:2.3.5]
>>>>>>> at
>>>>>>> org.glassfish.grizzly.filterchain.ExecutorResolver$9.excute(ExecutorRes
>>>>>>> olver.java:119) ~[grizzly-framework-2.3.5.ar:2.3.5]
>>>>>>> at
>>>>>>> org.glassfish.grizzly.filterchain.DefaultFilterChain.executeFilter(Defau
>>>>>>> ltFilterCain.java:288) ~[grizzly-framework-2.3.5.jar:2.3.5]
>>>>>>> at
>>>>>>> org.glassfish.grizzly.filterchain.efaultFilterChain.executeChainPart(De
>>>>>>> faultFilterChain.java:206) ~[grizzly-framework-2.3.5.jar:2.3.5]
>>>>>>> at
>>>>>>> org.glassfish.grizzly.filterchain.DefaultFilterChain.execute(DefaultFilt
>>>>>>> erChain.java:136) ~[grizzly-framework-2.3.5.jar:2.3.5]
>>>>>>> at
>>>>>>> org.glassfish.grizzly.filterchain.DefaultFilterChain.process(DefaultFilt
>>>>>>> erChain.java:11) ~[grizzly-framework-2.3.5.jar:2.3.5]
>>>>>>> at
>>>>>>> org.glassish.grizzly.ProcessorExecutor.execute(ProcessorExecutor.java:7
>>>>>>> 7) ~[grizzly-framework-2.3.5.jar:2.3.5]
>>>>>>> at
>>>>>>> org.glassfish.grizzly.nio.transport.TCPNIOTransport.fireIOEvent(TCPNIOTr
>>>>>>> ansport.java546) ~[grizzly-framework-2.3.5.jar:2.3.5]
>>>>>>> at
>>>>>>> org.glassfish.grizzly.strategies.AbstractIOStrategy.fireIOEvent(Abstract
>>>>>>> IOStrategy.java:113) ~[grizzly-framework-2.3..jar:2.3.5]
>>>>>>> at
>>>>>>> org.glassfish.grizzly.strategies.WorkerThreadIOStrategy.run0(WorkerThrea
>>>>>>> dIOStrategy.java:115) ~[grizzly-framework-2.3.5.jar:2.3.5]
>>>>>>> at
>>>>>>> org.glassfish.grizzly.strategies.WorkerThreadIOStrategy.access$100(Worke
>>>>>> rThreadIOStrategy.java:55) ~[grizzly-framework-2.3.5.jar:2.3.5]
>>>>>>> a
>>>>>>> org.glassfish.grizzly.strategies.WorkerThreadIOStrategy$WorkerThreadRunn
>>>>>>> able.run(WorkerThreadIOStrategy.java:135)
>>>>>>> ~[grizzl-framework-2.3.5.jar:2.3.5]
>>>>>>> at
>>>>>>> org.glassfish.grizzly.threadpool.AbstractThreadPool$Worker.doWork(Abstra
>>>>>>> ctThreadPool.java:565) ~[grizzly-framework-2.3.5.jar:2.3.5]
>>>>>>> at >>>>>>> org.glassfish.grizzly.threadpool.AbstractThreadPool$Worker.run(AbstractT
>>>>>>> hreadPool.java:545) ~[grizzly-framework-2.3.5.jar:2.3.5]
>>>>>>> Wrapped by: java.uti.concurrent.ExecutionException:
>>>>>>> java.io.IOException: Maximum pooled connections exceeded
>>>>>>> at
>>>>>>> org.glassfish.grizzly.impl.SafeFutureImpl$Syc.innerGet(SafeFutureImpl.j
>>>>>>> ava:359) ~[grizzly-framework-2.3.5.jar:2.3.5
>>>>>>> at
>>>>>>> org.glassfish.grizzly.impl.SafeFutureImpl.get(SafeFutureImpl.java:265)
>>>>>>> ~[grizzly-framework-2.3.5.jar:.3.5]
>>>>>>> at
>>>>>>> com.ning.http.client.providers.grizzly.GrizzlyResponseFuture.get(Grizzly
>>>>>>> ResponseFuture.java:165) ~[async-http-client-1.7.20-2846817.jar:na]
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>> boolean canReturnConnection(final Connection c) {
>>>>>>>
>>>>>>> return (DO_NOT_CACHE.get(c) != null ||
>>>>>>> pool.canCacheConnection());
>>>>>>>
>>>>>>> }
>>>>>>>
>>>>>>>
>>>>>>> private static HttpTransactionContext cleanup(final
>>>>>>> FilterChainContext ctx,
>>>>>>> final
>>>>>>> GrizzlyAsyncHttpProvider provider) {
>>>>>>>
>>>>>>> final Connection c = ctx.getConnectio();
>>>>>>> final HttpTransactionContext context =
>>>>>>> provider.getHttpTransactionContext(c);
>>>>>>> context.provider.setHttpTransactionContext(c,null);
>>>>>>> if
>>>>>>> (!context.provider.connectionManager.canReturnConnection(c)) {
>>>>>>> context.abort(new IOException("Maximum pooled
>>>>>>> connections exceeded"));
>>>>>>> } else {
>>>>>>> if
>>>>>>> (!context.provider.connectionManager.returnConnection(context.request,
>>>>>>> c)) {
>>>>>>> ctx.getConnection().close();
>>>>>>> }
>>>>>>> }
>>>>>>>
>>>>>>> return context;
>>>>>>>
>>>>>>> }
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>> We use:
>>>>>>>
>>>>>>> Key=[properties.httpclient.allowSslConnectionPool] Value=[true]
>>>>>>> Key=[properties.httpclient.maximumConnectionsPerHost] Value=[20]
>>>>>>> Key=[properties.httpclient.maximumConnectionsTotal] Value=[30]
>>>>>>> Key=[properties.httpclient.timeout.connection] Value=[10000]
>>>>>>> Key=[properties.httpclient.timeout.request] Value=[30000]
>>>>>>>
>>>>>>>
>>>>>>> In our logs, I can see there are 30 times this log:
>>>>>>>
>>>>>>> Remote closed connection
>>>>>>> (TCPNIOConnection{localSocketAddress={/172.16.104.160:55488
>>>>>>> <http://172.16.104.160:55488> },
>>>>>>> peerSocketAddress={host/172.16.4.100:443}}). Removing from cache
>>>>>>>
>>>>>>> and then it seems no connection is never available and the app is
>>>>>>> blocked.
>>>>>>>
>>>>>>> I'll try tomorrow to disable or increase the pool size, but if you have
>>>>>>> any idea of the problem please let me know.
>>>>>>> We do not run load tests, these are simple functional tests with nearly
>>>>>>> no concurrency.
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>> Thanks
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>> 2013/9/17 Ryan Lubke <ryan.lubke_at_oracle.com>
>>>>>>> This is good to hear. Thanks for all the feedback and working with us
>>>>>>> to nail this down.
>>>>>>>
>>>>>>>
>>>>>>> Sébastien Lorber wrote:
>>>>>>> Thanks,
>>>>>>>
>>>>>>> We already installed the previous snapshot in our nexus because it works
>>>>>>> fine and I'm working on something else now but I'll give you a feedback
>>>>>>> soon to see if this still works fine.
>>>>>>> For us it doesn't really mater which thread is using since we do not use
>>>>>>> Future in our applications for this case.
>>>>>>>
>>>>>>>
>>>>>>> Btw I've been able deploy a main with my code running concurrent
>>>>>>> uploads on a server which has a better connectivity with the remote API
>>>>>>> and it seems I can upload up to 150 concurrent files of 10mb with a heap
>>>>>>> of 250mo in 30 seconds and a throughput near 50mb/s
>>>>>>>
>>>>>>> I don't really know the infrastructure on which my code was deployed but
>>>>>>> I suspect Grizzly is not the bottleneck anymore :)
>>>>>>> When I have more than 150 concurrent uploads the remote endpoint seems
>>>>>>> to timeout under the load.
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>> 2013/9/16 Ryan Lubke <ryan.lubke_at_oracle.com>
>>>>>>>
>>>>>>>
>>>>>>> Ryan Lubke wrote:
>>>>>>>
>>>>>>>
>>>>>>> Sébastien Lorber wrote:
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>> I noticed something strange.
>>>>>>> On the FileInputStream I have, I've added a log on the close() of the
>>>>>>> stream which is called once the whole file has been read to be sent to
>>>>>>> the feeder.
>>>>>>>
>>>>>>> 1) If I perform a request (ie my session init request in the previous
>>>>>>> discussions) before doing my multipart upload, the thread that does
>>>>>>> execute the feeding is the thread that fires the request, and not a
>>>>>>> Grizzly worker.
>>>>>>> [Thread-30] INFO Will close file stream java.io.FileInputStream_at_27fe4315
>>>>>>> after having read 1440304 bytes
>>>>>>>
>>>>>>>
>>>>>>> 2) If I don't do any request before firing the multipart upload, the
>>>>>>> thread that does the feeding is a Grizzly threadpool worker thread:
>>>>>>> [Grizzly(22)] INFO Will close file stream
>>>>>>> java.io.FileInputStream_at_59ac4002 after having read 1440304 bytes
>>>>>>>
>>>>>>>
>>>>>>> Is this normal? I would expect a worker thread to always be used, and
>>>>>>> the main applicative thread that performs the request to never be
>>>>>>> bocking. (but it's not such an issue for me, we don't have a reactive
>>>>>>> on-blocking app anyway)
>>>>>>>
>>>>>>> It's currently expected behavior. Will need to re-evaluate this based
>>>>>>> on the new semantics of the FeedableBodyGenerator.
>>>>>>> I've committed a change for this. If, upon testing, you find there i
>>>>>>> still an issue, please let us know.
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>> On the 1st case, here's the stacktrace whenthe feed method is called:
>>>>>>>
>>>>>>> at
>>>>>>> com.ning.http.client.poviders.grizzly.FeedableBodyGenerator.initializeA
>>>>>>> synchronousTransfer(FeedableBodyGenerator.java:178)
>>>>>>> at
>>>>>>> com.ning.http.client.providers.grizzly.GrizzlyAsyncHttpPovider$BodyGene
>>>>>>> ratorBodyHandler.doHandle(GrizzlyAsyncHttpProvider.java:2210)
>>>>>>> at
>>>>>>> com.ning.httpclient.providers.grizzly.GrizzlyAsyncHttpProvider.sendRequ
>>>>>>> est(GrizzlyAsyncHttpProvider.java:564)
>>>>>>> at
>>>>>>> com.ning.http.client.providers.grizzly.GrizzlyAsyncHttpProvider$AsyncHtt
>>>>>>> pClientFilter.sendAsGrizzlyRequest(GrizzlyAsyncHttpProider.java:913)
>>>>>>> at
>>>>>>> com.ning.http.client.providers.grizzly.GrizzlyAsyncHttpProvider$AsyncHtt
>>>>>>> pClientFilter.handleWrite(GrizzlyAsyncHttpProvider.java:795)
>>>>>>> at
>>>>>>> org.glassfish.grizzly.filterchain.ExecutorResolver$8.execute(ExecutorRes
>>>>>>> olver.java:111)
>>>>>>> at
>>>>>>> org.glassfish.grizzly.filterchain.DefaultFilterChain.executeFilter(Defau
>>>>>>> ltFilterChain.java:288)
>>>>>>> at
>>>>>>> org.glassfish.grizzly.filterchain.DefaultFilterChain.executeChainPart(De
>>>>>>> faultFilterChain.java:206)
>>>>>>> at
>>>>>>> org.glassfish.grizzly.filterchain.DefaultFilterChain.execute(DefaultFilt
>>>>>>> erChain.java:136)
>>>>>>> at
>>>>>>> org.glassfish.grizzly.filterchain.DefaultFilterChain.process(DefaultFilt
>>>>>>> erChain.java:114)
>>>>>>> at
>>>>>>> org.glassfish.grizzly.ProcessrExecutor.execute(ProcessorExecutor.java:7
>>>>>>> 7)
>>>>>>> at
>>>>>>> org.lassfish.grizzly.filterchain.DefaultFilterChain.write(DefaultFilter
>>>>>>> Chain.java:437)
>>>>>>> at org.glassfish.grizzly.nio.NIOConnection.write(NIOConnection.java:387)
>>>>>>> at org.glassfish.grizzly.nio.NIOConnection.write(NIOConnection.java:361)
>>>>>>> at
>>>>>>> com.ning.http.client.providers.gizzly.GrizzlyAsyncHttpProvider.execute(
>>>>>>> GrizzlyAsyncHttpProvider.java307)
>>>>>>> at
>>>>>>> com.ning.http.client.providers.grizzly.GrizzlyAsyncHttpProvider$1.comple
>>>>>>>ted(GrizzlyAsyncHttpProvider.java:224)
>>>>>>> at
>>>>>>> com.ning.http.clientproviders.grizzly.GrizzlyAsyncHttpProvider$1.comple
>>>>>>> ted(GrizzlyAsyncHttpProvider.java:210)
>>>>>>> at
>>>>>>> com.nng.http.client.providers.grizzly.GrizzlyAsyncHttpProvider$Connecti
>>>>>>> onManager.doAsyncTrackedConnection(GrizzlyAsyncHttpProvider.java:289)
>>>>>>> at
>>>>>>> com.ning.http.client.providers.grizzly.GrizzlyAsyncHttpProvider.execute(
>>>>>>> GrizzlyAsyncHttpProvider.java:244)
>>>>>>> at
>>>>>>> com.ning.http.client.AsyncHttpClient.executeRequest(AsyncHttpClient.java
>>>>>>> :534)
>>>>>>>
>>>>>>> So it seems this happen when the handshake has already been done when
>>>>>>> initializeAsynchronousTransfer is called, so that we do not go through
>>>>>>> the HandshakeListener
>>>>>>
>>>>>>>
>>>>>>> Notice that this case seems to also happen in a few threads on the 2nd
>>>>>>> case, but most of the threads are Grizzly workers.
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>> 2013/9/12 Sébastien Lorber <lober.sebastien_at_gmail.com>
>>>>>>>
>>>>>>>
>>>>>>> Hi.
>>>>>>>
>>>>>>>
>>>>>> Thanks, it seems to work.
>>>>>>>
>>>>>>> I would suggest to throw IOException on the flush method since the feed
>>>>>>> method is supposed to be called here
>>>>>>>
>>>>>>>
>>>>>>> My implementation of flush is:
>>>>>>>
>>>>>> @Override
>>>>>>> public void flush() {
>>>>>>> Part[] partsArray = arts.toArray(new Part[parts.size()]);
>>>>>>> try ( OutputStream outputSream = createFeedingOutputStream() ) {
>>>>>>> Part.sendParts(outputStream,partsArray,multipartBoundary);
>>>>>>> } catch (Exception e) {
>>>>>>> throw new IllegalStateException("Unable to feed the
>>>>>>> FeedableBodyGenerator",e);
>>>>>>> }
>>>>>>> }
>>>>>>>
>>>>>>> Is this correct? The OutputStream redirects the bytes written to the
>>>>>>> feed(Buffer) method
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>> There seems to be some concurrency issue. Because the upload of 1 file
>>>>>>> seems fine, but when using multiple threads, I often get the following
>>>>>>> stack:
>>>>>>> Caused by: java.io.IOException: Stream Closed
>>>>>>> at java.io.FileInputStream.readBytes(Native Method)
>>>>>>> at java.io.FileInputStream.read(FileInputStream.java:242)
>>>>>>> at
>>>>>>> com.google.common.io.CountingInputStream.read(CountingInputStream.java:6
>>>>>>> 2)
>>>>>>> at java.io.FilterInputStream.read(FilterInputStream.java:133)
>>>>>>> at java.io.FilterInputStream.read(FilterInputStream.java:107)
>>>>>>> at com.ning.http.multipart.FilePart.sendData(FilePart.java:178)
>>>>>>> at com.ning.http.multipart.Part.send(Part.java:331)
>>>>>>> at com.ning.http.multipart.Part.sendParts(Part.java:397)
>>>>>>>
>>>>>>>
>>>>>>> This is because the flush() method is called multiple times for the same
>>>>>>> request on some cases.
>>>>>>> I guess this is not supposed to happen.
>>>>>>> What I understand is that the flush() method is supposed to be called
>>>>>>> only once.
>>>>>>>
>>>>>>>
>>>>>>> Using debug logging breakpoints I get the following:
>>>>>>>
>>>>>>> myapp--api-test 12/09/2013-12:20:46.042 [] [] [main] INFO
>>>>>>> com.myapp.perf.DocumentUploadPerfIntegrationTest:77 - Thread 0 started
>>>>>>> myapp--api-test 12/09/2013-12:20:46.042 [] [] [main] INFO
>>>>>>> com.myapp.perf.DocumentUploadPerfIntegrationTest:77 - Thread 1 started
>>>>>>> myapp--api-test 12/09/2013-12:20:46.043 [] [] [main] INFO
>>>>>>> com.myapp.perf.DocumentUploadPerfIntegrationTest:77 - Thread 2 started
>>>>>>> myapp--api-test 12/09/2013-12:20:46.043 [] [] [main] INFO
>>>>>>> com.myapp.perf.DocumentUploadPerfIntegrationTest:77 - Thread 3 started
>>>>>>> myapp--api-test 12/09/2013-12:20:46.043 [] [] [main] INFO
>>>>>>> com.myapp.perf.DocumentUploadPerfIntegrationTest:77 - Thread 4 started
>>>>>>> myapp--api-test 12/09/2013-12:20:46.044 [] [] [main] INFO
>>>>>>> com.myapp.perf.DocumentUploadPerfIntegrationTest:77 - Thread 5 started
>>>>>>> myapp--api-test 12/09/2013-12:20:46.044 [] [] [main] INFO
>>>>>>> com.myapp.perf.DocumentUploadPerfIntegrationTest:77 - Thread 6 started
>>>>>>> myapp--api-test 12/09/2013-12:20:46.044 [] [] [main] INFO
>>>>>>> com.myapp.perf.DocumentUploadPerfIntegrationTest:77 - Thread 7 started
>>>>>>> myapp--api-test 12/09/2013-12:20:46.045 [] [] [main] INFO
>>>>>>> com.myapp.perf.DocumentUploadPerfIntegrationTest:77 - Thread 8 started
>>>>>>> myapp--api-test 12/09/2013-12:20:46.045 [] [] [main] INFO
>>>>>>> com.myapp.perf.DocumentUploadPerfIntegrationTest:77 - Thread 9 started
>>>>>>> myapp--api-test 12/09/2013-12:20:46.045 [] [] [main] INFO
>>>>>>> com.myapp.perf.DocumentUploadPerfIntegrationTest:77 - Thread 10 started
>>>>>>> myapp--api-test 12/09/2013-12:20:46.046 [] [] [main] INFO
>>>>>>> com.myapp.perf.DocumentUploadPerfIntegrationTest:77 - Thread 11 started
>>>>>>> myapp--api-test 12/09/2013-12:20:46.047 [] [] [main] INFO
>>>>>>> com.myapp.perf.DocumentUploadPerfIntegrationTest:77 - Thread 12 started
>>>>>>> myapp--api-test 12/09/2013-12:20:46.048 [] [] [main] INFO
>>>>>>> com.myapp.perf.DocumentUploadPerfIntegrationTest:77 - Thread 13 started
>>>>>>> myapp--api-test 12/09/2013-12:20:46.049 [] [] [main] INFO
>>>>>>> com.myapp.perf.DocumentUploadPerfIntegrationTest:77 - Thread 14 started
>>>>>>> myapp--api-test 12/09/2013-12:20:46.049 [] [] [main] INFO
>>>>>>> com.myapp.perf.DocumentUploadPerfIntegrationTest:77 - Thread 15 started
>>>>>>> myapp--api-test 12/09/2013-12:20:46.050 [] [] [main] INFO
>>>>>>> com.myapp.perf.DocumentUploadPerfIntegrationTest:77 - Thread 16 started
>>>>>>> myapp--api-test 12/09/2013-12:20:46.050 [] [] [main] INFO
>>>>>>> com.myapp.perf.DocumentUploadPerfIntegrationTest:77 - Thread 17 started
>>>>>>> myapp--api-test 12/09/2013-12:20:46.051 [] [] [main] INFO
>>>>>>> com.myapp.perf.DocumentUploadPerfIntegrationTest:77 - Thread 18 started
>>>>>>> myapp--api-test 12/09/2013-12:20:46.051 [] [] [main] INFO
>>>>>>> com.myapp.perf.DocumentUploadPerfIntegrationTest:77 - Thread 19 started
>>>>>>> Adding handshake listener for
>>>>>>> com.ning.http.client.providers.grizzly.FeedableBodyGenerator_at_417de6ff
>>>>>>> Adding handshake listener for
>>>>>>> com.ning.http.client.providers.grizzly.FeedableBodyGenerator_at_6c0d6ef7
>>>>>>> Adding handshake listener for
>>>>>>> com.ning.http.client.providers.grizzly.FeedableBodyGenerator_at_7799b411
>>>>>>> Adding handshake listener for
>>>>>>> com.ning.http.client.providers.grizzly.FeedableBodyGenerator_at_1c940409
>>>>>>> Adding handshake listener for
>>>>>>> com.ning.http.client.providers.grizzly.FeedableBodyGenerator_at_480f9510
>>>>>>> Adding handshake listener for
>>>>>>> com.ning.http.client.providers.grizzly.FeedableBodyGenerator_at_3e888183
>>>>>>> Adding handshake listener for
>>>>>>> com.ning.http.client.providers.grizzly.FeedableBodyGenerator_at_17840db8
>>>>>>> Adding handshake listener for
>>>>>>> com.ning.http.client.providers.grizzly.FeedableBodyGenerator_at_2cbad94b
>>>>>>> Adding handshake listener for
>>>>>>> com.ning.http.client.providers.grizzly.FeedableBodyGenerator_at_64c0a4ae
>>>>>>> Adding handshake listener for
>>>>>>> com.ning.http.client.providers.grizzly.FeedableBodyGenerator_at_102873a6
>>>>>>> Adding handshake listener for
>>>>>>> com.ning.http.client.providers.grizzly.FeedableBodyGenerator_at_3d5d9ee8
>>>>>>> Adding handshake listener for
>>>>>>> com.ning.http.client.providers.grizzly.FeedableBodyGenerator_at_51557949
>>>>>>> Adding handshake listener for
>>>>>>> com.ning.http.client.providers.grizzly.FeedableBodyGenerator_at_6f95de2f
>>>>>>> Adding handshake listener for
>>>>>>> com.ning.http.client.providers.grizzly.FeedableBodyGenerator_at_346d784c
>>>>>>> Completing and removing handshake listener for
>>>>>>> com.ning.http.client.providers.grizzly.FeedableBodyGenerator_at_7799b411
>>>>>>> Completing and removing handshake listener for
>>>>>>> com.ning.http.client.providers.grizzly.FeedableBodyGenerator_at_5befaa07
>>>>>>> Completing and removing handshake listener for
>>>>>>> com.ning.http.client.providers.grizzly.FeedableBodyGenerator_at_5befaa07
>>>>>>> Completing and removing handshake listener for
>>>>>>> com.ning.http.client.providers.grizzly.FeedableBodyGenerator_at_5befaa07
>>>>>>> Completing and removing handshake listener for
>>>>>>> com.ning.http.client.providers.grizzly.FeedableBodyGenerator_at_102873a6
>>>>>>> Completing and removing handshake listener for
>>>>>>> com.ning.http.client.providers.grizzly.FeedableBodyGenerator_at_102873a6
>>>>>>> Completing and removing handshake listener for
>>>>>>> com.ning.http.client.providers.grizzly.FeedableBodyGenerator_at_346d784c
>>>>>>> Completing and removing handshake listener for
>>>>>>> com.ning.http.client.providers.grizzly.FeedableBodyGenerator_at_64c0a4ae
>>>>>>> Completing and removing handshake listener for
>>>>>>> com.ning.http.client.providers.grizzly.FeedableBodyGenerator_at_674735a8
>>>>>>> Completing and removing handshake listener for
>>>>>>> com.ning.http.client.providers.grizzly.FeedableBodyGenerator_at_64c0a4ae
>>>>>>> Completing and removing handshake listener for
>>>>>>> com.ning.http.client.providers.grizzly.FeedableBodyGenerator_at_480f9510
>>>>>>> Completing and removing handshake listener for
>>>>>>> com.ning.http.client.providers.grizzly.FeedableBodyGenerator_at_6c0d6ef7
>>>>>>> Completing and removing handshake listener for
>>>>>>> com.ning.http.client.providers.grizzly.FeedableBodyGenerator_at_1daf8fd8
>>>>>>>
>>>>>>>
>>>>>>> As you can see the same HandshakeListener seems to be called multiple
>>>>>>> times for the same FeedableBodyGenerator
>>>>>>>
>>>>>>> When this happens, the stack is:
>>>>>>>
>>>>>>> at
>>>>>>> com.ning.http.client.providers.grizzly.FeedableBodyGenerator$1.onComplet
>>>>>>> e(FeedableBodyGenerator.java:198)
>>>>>>> ~[async-http-client-1.7.20-SNAPSHOT.jar:na]
>>>>>>> at
>>>>>>> org.glassfish.grizzly.ssl.SSLBaseFilter.notifyHandshakeComplete(SSLBaseF
>>>>>>> ilter.java:880) ~[grizzly-framework-2.3.5.jar:2.3.5]
>>>>>>> at
>>>>>>> org.glassfish.grizzly.ssl.SSLFilter.notifyHandshakeComplete(SSLFilter.ja
>>>>>>> va:282) ~[grizzly-framework-2.3.5.jar:2.3.5]
>>>>>>> at
>>>>>>> org.glassfish.grizzly.ssl.SSLBaseFilter.handleRead(SSLBaseFilter.java:27
>>>>>>> 5) ~[grizzly-framework-2.3.5.jar:2.3.5]
>>>>>>> at
>>>>>>> com.ning.http.client.providers.grizzly.GrizzlyAsyncHttpProvider$Switchin
>>>>>>> gSSLFilter.handleRead(GrizzlyAsyncHttpProvider.java:2490)
>>>>>>> ~[async-http-client-1.7.20-SNAPSHOT.jar:na]
>>>>>>> at
>>>>>>> org.glassfish.grizzly.filterchain.ExecutorResolver$9.execute(ExecutorRes
>>>>>>> olver.java:119) ~[grizzly-framework-2.3.5.jar:2.3.5]
>>>>>>> at
>>>>>>> org.glassfish.grizzly.filterchain.DefaultFilterChain.executeFilter(Defau
>>>>>>> ltFilterChain.java:288) ~[grizzly-framework-2.3.5.jar:2.3.5]
>>>>>>> at
>>>>>>> org.glassfish.grizzly.filterchain.DefaultFilterChain.executeChainPart(De
>>>>>>> faultFilterChain.java:206) ~[grizzly-framework-2.3.5.jar:2.3.5]
>>>>>>> at
>>>>>>> org.glassfish.grizzly.filterchain.DefaultFilterChain.execute(DefaultFilt
>>>>>>> erChain.java:136) ~[grizzly-framework-2.3.5.jar:2.3.5]
>>>>>>> at
>>>>>>> org.glassfish.grizzly.filterchain.DefaultFilterChain.process(DefaultFilt
>>>>>>> erChain.java:114) ~[grizzly-framework-2.3.5.jar:2.3.5]
>>>>>>> at
>>>>>>> org.glassfish.grizzly.ProcessorExecutor.execute(ProcessorExecutor.java:7
>>>>>>> 7) ~[grizzly-framework-2.3.5.jar:2.3.5]
>>>>>>> at
>>>>>>> org.glassfish.grizzly.nio.transport.TCPNIOTransport.fireIOEvent(TCPNIOTr
>>>>>>> ansport.java:546) ~[grizzly-framework-2.3.5.jar:2.3.5]
>>>>>>> at
>>>>>>> org.glassfish.grizzly.strategies.AbstractIOStrategy.fireIOEvent(Abstract
>>>>>>> IOStrategy.java:113) ~[grizzly-framework-2.3.5.jar:2.3.5]
>>>>>>> at
>>>>>>> org.glassfish.grizzly.strategies.WorkerThreadIOStrategy.run0(WorkerThrea
>>>>>>> dIOStrategy.java:115) ~[grizzly-framework-2.3.5.jar:2.3.5]
>>>>>>> at
>>>>>>> org.glassfish.grizzly.strategies.WorkerThreadIOStrategy.access$100(Worke
>>>>>>> rThreadIOStrategy.java:55) ~[grizzly-framework-2.3.5.jar:2.3.5]
>>>>>>> at
>>>>>>> org.glassfish.grizzly.strategies.WorkerThreadIOStrategy$WorkerThreadRunn
>>>>>>> able.run(WorkerThreadIOStrategy.java:135)
>>>>>>> ~[grizzly-framework-2.3.5.jar:2.3.5]
>>>>>>> at
>>>>>>> org.glassfish.grizzly.threadpool.AbstractThreadPool$Worker.doWork(Abstra
>>>>>>> ctThreadPool.java:565) ~[grizzly-framework-2.3.5.jar:2.3.5]
>>>>>>> at
>>>>>>> org.glassfish.grizzly.threadpool.AbstractThreadPool$Worker.run(AbstractT
>>>>>>> hreadPool.java:545) ~[grizzly-framework-2.3.5.jar:2.3.5]
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>> So I tried with the following code:
>>>>>>>
>>>>>>> private boolean alreadyFlushed = false;
>>>>>>> @Override
>>>>>>> public synchronized void flush() {
>>>>>>> if ( alreadyFlushed ) {
>>>>>>> return;
>>>>>>> }
>>>>>>> startFeeding();
>>>>>>> alreadyFlushed = true;
>>>>>>> }
>>>>>>>
>>>>>>> It works fine when upload small files.
>>>>>>>
>>>>>>> But with larger files, I often get TimeoutException stacks for some of
>>>>>>> the threads:
>>>>>>>
>>>>>>> Caused by: java.util.concurrent.TimeoutException
>>>>>>> at
>>>>>>> org.glassfish.grizzly.impl.SafeFutureImpl$Sync.innerGet(SafeFutureImpl.j
>>>>>>> ava:367)
>>>>>>> at
>>>>>>> org.glassfish.grizzly.impl.SafeFutureImpl.get(SafeFutureImpl.java:274)
>>>>>>> at
>>>>>>> com.ning.http.client.providers.grizzly.FeedableBodyGenerator$BaseFeeder.
>>>>>>> block(FeedableBodyGenerator.java:349)
>>>>>>> at
>>>>>>> com.ning.http.client.providers.grizzly.FeedableBodyGenerator$BaseFeeder.
>>>>>>> blockUntilQueueFree(FeedableBodyGenerator.java:339)
>>>>>>> at
>>>>>>> com.ning.http.client.providers.grizzly.FeedableBodyGenerator$BaseFeeder.
>>>>>>> feed(FeedableBodyGenerator.java:306)
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>> Did I miss something?
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>> 2013/9/12 Ryan Lubke <ryan.lubke_at_oracle.com>
>>>>>>> Committed another small change. Please make sure you're at the latest
>>>>>>> when you build.
>>>>>>>
>>>>>>> -rl
>>>>>>>
>>>>>>>
>>>>>>> Ryan Lubke wrote:
>>>>>>> Okay, I've committed another set of refactorings to the
>>>>>>> FeedableBodyGenerator.
>>>>>>>
>>>>>>> For your use case, you should extend FeedableBodyGenerator.SimpleFeeder.
>>>>>>>
>>>>>>> Let me know if you run into issues.
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>> Sébastien Lorber wrote:
>>>>>>> Yes I think it would, so that I can feed the queue at once
>>>>>>>
>>>>>>> One thread will be locked during the feeding for nothing but it's not a
>>>>>>> real problem in my usecase.
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>> 2013/9/10 Ryan Lubke <ryan.lubke_at_oracle.com>
>>>>>>> Would having a different listener that will be notified once async
>>>>>>> transferring has been started work better for you?
>>>>>>>
>>>>>>> Something like:
>>>>>>>
>>>>>>> onAsyncTransferInitiated() {
>>>>>>> // invoke your feed method
>>>>>>> }
>>>>>>>
>>>>>>> ?
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>> Sébastien Lorber wrote:
>>>>>>> Unfortunatly I won't be able to use the Feeder non-blocking stuff for
>>>>>>> now, because of how the multipart request in handled in AHC
>>>>>>>
>>>>>>>
>>>>>>> Here's my feeding method:
>>>>>>>
>>>>>>> public void feed() throws IOException {
>>>>>>> Part[] partsArray = parts.toArray(new Part[parts.size()]);
>>>>>>> try ( OutputStream outputStream = createFeedingOutputStream() ) {
>>>>>>> Part.sendParts(outputStream,partsArray,multipartBoundary);
>>>>>>> } catch (Exception e) {
>>>>>>> throw new IllegalStateException("Unable to feed the
>>>>>>> FeedableBodyGenerator",e);
>>>>>>> }
>>>>>>> }
>>>>>>>
>>>>>>>
>>>>>>> As you can see, the multipart Parts array can only be pushed to the
>>>>>>> OutputStream, I don't have any way to "pull" the data when the canFeed()
>>>>>>> method is triggered.
>>>>>>>
>>>>>>>
>>>>>>> But I've seen that there's a com.ning.http.multipart.MultipartBody#read
>>>>>>> that seems to provide a memory efficient way to pull data from a
>>>>>>> Multipart body...
>>>>>>>
>>>>>>> Should see what I come up with this
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>> 2013/9/10 Sébastien Lorber <lorber.sebastien_at_gmail.com>
>>>>>>> It seems the Feeder is highly recommended but not mandatory so I tried
>>>>>>> without.
>>>>>>>
>>>>>>> With my existing code it seems there is a synchronization problem.
>>>>>>>
>>>>>>>
>>>>>>> The feeding threads get locked to the prematureFeed.get();
>>>>>>>
>>>>>>> So the Grizzly kernel threads are unable to acquire the lock required to
>>>>>>> enter the initializeAsynchronousTransfer method
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>> Will try with an implementation of Feeder
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>> 2013/9/10 Sébastien Lorber <lorber.sebastien_at_gmail.com>
>>>>>>> Hmmm it seems I have a problem with one of your maven plugins. I'll try
>>>>>>> to bypass it, but for info:
>>>>>>>
>>>>>>> ➜ ahc2 git:(ahc-1.7.x) mvn clean install
>>>>>>> [WARNING]
>>>>>>> [WARNING] Some problems were encountered while building the effective
>>>>>>> settings
>>>>>>> [WARNING] 'profiles.profile[default].repositories.repository.id
>>>>>>> <http://repositories.repository.id> ' must be unique but found duplicate
>>>>>>> repository with id fullsix-maven-repository @
>>>>>>> /home/slorber/.m2/settings.xml
>>>>>>> [WARNING]
>>>>>>> [INFO] Scanning for projects...
>>>>>>> [INFO]
>>>>>>> [INFO]
>>>>>>> ------------------------------------------------------------------------
>>>>>>> [INFO] Building Asynchronous Http Client 1.7.20-SNAPSHOT
>>>>>>> [INFO]
>>>>>>> ------------------------------------------------------------------------
>>>>>>> [INFO]
>>>>>>> [INFO] --- maven-clean-plugin:2.4.1:clean (default-clean) @
>>>>>>> async-http-client ---
>>>>>>> [INFO]
>>>>>>> [INFO] --- maven-enforcer-plugin:1.0-beta-1:enforce (enforce-maven) @
>>>>>>> async-http-client ---
>>>>>>> [INFO]
>>>>>>> [INFO] --- maven-enforcer-plugin:1.0-beta-1:enforce (enforce-versions) @
>>>>>>> async-http-client ---
>>>>>>> [INFO]
>>>>>>> [INFO] --- maven-resources-plugin:2.4.3:resources (default-resources) @
>>>>>>> async-http-client ---
>>>>>>> [INFO] Using 'UTF-8' encoding to copy filtered resources.
>>>>>>> [INFO] skip non existing resourceDirectory
>>>>>>> /home/slorber/Bureau/ahc2/src/main/resources
>>>>>>> [INFO]
>>>>>>> [INFO] --- maven-compiler-plugin:2.3.2:compile (default-compile) @
>>>>>>> async-http-client ---
>>>>>>> [INFO] Compiling 158 source files to
>>>>>>> /home/slorber/Bureau/ahc2/target/classes
>>>>>>> [INFO]
>>>>>>> [INFO] --- animal-sniffer-maven-plugin:1.6:check (check-java-1.5-compat)
>>>>>>> @ async-http-client ---
>>>>>>> [INFO] Checking unresolved references to
>>>>>>> org.codehaus.mojo.signature:java15:1.0
>>>>>>> [ERROR] Undefined reference:
>>>>>>> java/io/IOException.<init>(Ljava/lang/Throwable;)V in
>>>>>>> /home/slorber/Bureau/ahc2/target/classes/com/ning/http/client/providers/
>>>>>>> grizzly/FeedableBodyGenerator.class
>>>>>>> [INFO]
>>>>>>> ------------------------------------------------------------------------
>>>>>>> [INFO] BUILD FAILURE
>>>>>>> [INFO]
>>>>>>> ------------------------------------------------------------------------
>>>>>>> [INFO] Total time: 8.747s
>>>>>>> [INFO] Finished at: Tue Sep 10 11:25:41 CEST 2013
>>>>>>> [INFO] Final Memory: 30M/453M
>>>>>>> [INFO]
>>>>>>> ------------------------------------------------------------------------
>>>>>>> [ERROR] Failed to execute goal
>>>>>>> org.codehaus.mojo:animal-sniffer-maven-plugin:1.6:check
>>>>>>> (check-java-1.5-compat) on project async-http-client: Signature errors
>>>>>>> found. Verify them and put @IgnoreJRERequirement on them. -> [Help 1]
>>>>>>> [ERROR]
>>>>>>> [ERROR] To see the full stack trace of the errors, re-run Maven with the
>>>>>>> -e switch.
>>>>>>> [ERROR] Re-run Maven using the -X switch to enable full debug logging.
>>>>>>> [ERROR]
>>>>>>> [ERROR] For more information about the errors and possible solutions,
>>>>>>> please read the following articles:
>>>>>>> [ERROR] [Help 1]
>>>>>>> http://cwiki.apache.org/confluence/display/MAVEN/MojoFailureException
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>> 2013/9/10 Sébastien Lorber <lorber.sebastien_at_gmail.com>
>>>>>>> Ok thank you, I'll try to implement that today and will give you my
>>>>>>> feedback :)
>>>>>>>
>>>>>>>
>>>>>>> 2013/9/10 Ryan Lubke <ryan.lubke_at_oracle.com>
>>>>>>> Okay,
>>>>>>>
>>>>>>> I've committed my initial changes to the AHC repository. Here's a
>>>>>>> summary of the changes:
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>> Improvements to the FeedableBodyGenerator (Grizzly's).
>>>>>>> - Don't allow queueing of data before initiateAsyncTransfer has been
>>>>>>> invoked. In low memory
>>>>>>> heaps, this could lead to an OOM if the source is feeding too fast. The
>>>>>>> new behavior is to
>>>>>>> block until initiateAsyncTransfer is called, at which time the blocked
>>>>>>> thread may proceed with
>>>>>>> the feed operation.
>>>>>>> - Introduce the concept of a Feeder. Implementations are responsible,
>>>>>>> at a high level, for:
>>>>>>> + letting the provider know that data is available to be fed without
>>>>>>> blocking
>>>>>>> + allowing the registration of a callback that the Feeder
>>>>>>> implementation may invoke
>>>>>>> to signal that more data is available, if it wasn't available at a
>>>>>>> previous point in time.
>>>>>>> - When using a Feeder with a secure request, the SSL handshake will be
>>>>>>> kicked off by the
>>>>>>> initiateAsyncTransfer call, but feeding of data will not occur until
>>>>>>> the handshake is complete.
>>>>>>> This is necessary as the SSLFilter will queue up all writes until the
>>>>>>> handshake is complete,
>>>>>>> and currently, the buffer isn't tied in with the transport flow control
>>>>>>> mechanism.
>>>>>>> NOTE: This new SSL behavior is not currently applied when invoking the
>>>>>>> feed() method
>>>>>>> outside the context of a Feeder. Still need to address that.
>>>>>>> - Exposed configuration of the async write queue limit through the
>>>>>>> FeedableBodyGenerator.
>>>>>>> This is an improvement on using a TransportCustomizer as any
>>>>>>> configuration there is
>>>>>>> transport-wide, and therefor applied to all Connections. By exposing
>>>>>>> it here, each feeder
>>>>>>> may have a different byte limit.
>>>>>>> - Improved documentation for this class
>>>>>>>
>>>>>>> I recommend reading through the javadoc comments in the source [1] for
>>>>>>> FeedableBodyGenerator (comments welcome).
>>>>>>> Additionally, I would re-work your code to leverage the Feeder instead
>>>>>>> of calling feed() directly.
>>>>>>>
>>>>>>> If you have issues implementing Feeder, do let us know.
>>>>>>>
>>>>>>> If you have additional questions, again, let us know.
>>>>>>>
>>>>>>> Thanks,
>>>>>>> -rl
>>>>>>>
>>>>>>> [1]
>>>>>>> https://github.com/AsyncHttpClient/async-http-client/blob/ahc-1.7.x/src/
>>>>>>> main/java/com/ning/http/client/providers/grizzly/FeedableBodyGenerator.j
>>>>>>> ava
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>> Ryan Lubke wrote:
>>>>>>>
>>>>>>>
>>>>>>> Sébastien Lorber wrote:
>>>>>>> So in the end I've end up with an implementation that's working for me.
>>>>>>>
>>>>>>>
>>>>>>> I think there are 2 bugs:
>>>>>>>
>>>>>>> 1) The bytes can accumulate in the FeedableBodyGenerator queue if the
>>>>>>> initialize(ctx) method is not called fast enough.
>>>>>>> This can be solved by using a BlockingQueue of size 1 and the put()
>>>>>>> method.
>>>>>>>
>>>>>>> 2) Once the context is injected, the FeedableBodyGenerator flushes the
>>>>>>> queue.
>>>>>>> The matter is that if the connection is new, not warmed up by a
>>>>>>> previous request, then the SSL handshake is not done yet, and it seems
>>>>>>> that the bytes are accumulated in some part of the SSL filter which
>>>>>>> doesn't deliver them to the connection until the handshake has
>>>>>>> completed, so c.canWrite() continues to return true.
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>> I have replaced some part of the FeedableBodyGenerator to test this and
>>>>>>> it works pretty fine. See what I have changed:
>>>>>>>
>>>>>>> 1)
>>>>>>> private final BlockingQueue<BodyPart> queue = new
>>>>>>> LinkedBlockingQueue<BodyPart>(1);
>>>>>>>
>>>>>>>
>>>>>>> 2)
>>>>>>> public void feed(final Buffer buffer, final boolean last) throws
>>>>>>> IOException {
>>>>>>> try {
>>>>>>> queue.put(new BodyPart(buffer, last));
>>>>>>> } catch (InterruptedException e) {
>>>>>>> throw new RuntimeException(e);
>>>>>>> }
>>>>>>> queueSize.incrementAndGet();
>>>>>>>
>>>>>>> if (context != null) {
>>>>>>> blockUntilConnectionIsReadyToWrite(context);
>>>>>>> flushQueue(true);
>>>>>>> }
>>>>>>> }
>>>>>>>
>>>>>>> private void blockUntilConnectionIsReadyToWrite(FilterChainContext
>>>>>>> fcc) {
>>>>>>> if ( !connectionIsReadyToWrite(fcc) ) {
>>>>>>> while ( !connectionIsReadyToWrite(fcc) ) {
>>>>>>> try { Thread.sleep(10); } catch ( Exception e ) { throw new
>>>>>>> RuntimeException(e); }
>>>>>>> }
>>>>>>> }
>>>>>>> }
>>>>>>>
>>>>>>> private boolean connectionIsReadyToWrite(FilterChainContext fcc) {
>>>>>>> Connection connection = fcc.getConnection();
>>>>>>> SSLEngine sslEngine = SSLUtils.getSSLEngine(connection);
>>>>>>> return sslEngine != null && !SSLUtils.isHandshaking(sslEngine);
>>>>>>> }
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>> What do you think?
>>>>>>>
>>>>>>>
>>>>>>> We had come to similar conclusions on this end. I'm still working
>>>>>>> through testing the idea I mentioned previously (took longer than I
>>>>>>> expected - sorry).
>>>>>>> I hope to have something for you to test very soon.
>>>>>>>
>>>>>>> Note that it will be taking the above into account as well.
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>> 2013/9/5 Sébastien Lorber <lorber.sebastien_at_gmail.com>
>>>>>>>
>>>>>>> I have tried to put a while ( context == null ) Thread.sleep but it
>>>>>>> doesn't seem to work, when the context gets injected, after the sleeps,
>>>>>>> there's an OOM
>>>>>>>
>>>>>>> So I hope you'll have more success with your alternative :)
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>> I have done another test, remember my code that worked, which previously
>>>>>>> "warmed" the Thread with an useless request.
>>>>>>>
>>>>>>> private Runnable uploadSingleDocumentRunnable = new Runnable() {
>>>>>>> @Override
>>>>>>> public void run() {
>>>>>>> try {
>>>>>>> getUselessSessionCode();
>>>>>>> Thread.sleep(X);
>>>>>>> uploadSingleDocument();
>>>>>>> } catch ( Exception e ) {
>>>>>>> throw new RuntimeException("file upload failed",e);
>>>>>>> }
>>>>>>> }
>>>>>>> };
>>>>>>>
>>>>>>> I have put a sleep of X between the useless warmup request, and the real
>>>>>>> upload request
>>>>>>>
>>>>>>>
>>>>>>> What I noticed is that there is a very different behavior according to
>>>>>>> the value of X
>>>>>>>
>>>>>>>
>>>>>>> Under 10 seconds, it seems the stuff is still warm, I can upload the
>>>>>>> documents.
>>>>>>> Around 10 seconds I get a stack which seems to be "connection closed" or
>>>>>>> something
>>>>>>> Above 10 seconds, I get OOM like if the stuff wasn't warm.
>>>>>>>
>>>>>>>
>>>>>>> The stacks I get for 10 seconds looks like
>>>>>>>
>>>>>>> Caused by: javax.net.ssl.SSLException: SSLEngine is CLOSED
>>>>>>> at
>>>>>>> org.glassfish.grizzly.ssl.SSLConnectionContext.wrap(SSLConnectionContext
>>>>>>> .java:295) ~[grizzly-framework-2.3.5.jar:2.3.5]
>>>>>>> at
>>>>>>> org.glassfish.grizzly.ssl.SSLConnectionContext.wrapAll(SSLConnectionCont
>>>>>>> ext.java:238) ~[grizzly-framework-2.3.5.jar:2.3.5]
>>>>>>> at
>>>>>>> org.glassfish.grizzly.ssl.SSLBaseFilter.wrapAll(SSLBaseFilter.java:405)
>>>>>>> ~[grizzly-framework-2.3.5.jar:2.3.5]
>>>>>>> at
>>>>>>> org.glassfish.grizzly.ssl.SSLBaseFilter.handleWrite(SSLBaseFilter.java:3
>>>>>>> 20) ~[grizzly-framework-2.3.5.jar:2.3.5]
>>>>>>> at org.glassfish.grizzly.ssl.SSLFilter.accurateWrite(SSLFilter.java:255)
>>>>>>> ~[grizzly-framework-2.3.5.jar:2.3.5]
>>>>>>> at org.glassfish.grizzly.ssl.SSLFilter.handleWrite(SSLFilter.java:143)
>>>>>>> ~[grizzly-framework-2.3.5.jar:2.3.5]
>>>>>>> at
>>>>>>> com.ning.http.client.providers.grizzly.GrizzlyAsyncHttpProvider$Switchin
>>>>>>> gSSLFilter.handleWrite(GrizzlyAsyncHttpProvider.java:2500)
>>>>>>> ~[async-http-client-1.7.20-SNAPSHOT.jar:na]
>>>>>>> at
>>>>>>> org.glassfish.grizzly.filterchain.ExecutorResolver$8.execute(ExecutorRes
>>>>>>> olver.java:111) ~[grizzly-framework-2.3.5.jar:2.3.5]
>>>>>>> at
>>>>>>> org.glassfish.grizzly.filterchain.DefaultFilterChain.executeFilter(Defau
>>>>>>> ltFilterChain.java:288) ~[grizzly-framework-2.3.5.jar:2.3.5]
>>>>>>> at
>>>>>>> org.glassfish.grizzly.filterchain.DefaultFilterChain.executeChainPart(De
>>>>>>> faultFilterChain.java:206) ~[grizzly-framework-2.3.5.jar:2.3.5]
>>>>>>> at
>>>>>>> org.glassfish.grizzly.filterchain.DefaultFilterChain.execute(DefaultFilt
>>>>>>> erChain.java:136) ~[grizzly-framework-2.3.5.jar:2.3.5]
>>>>>>> at
>>>>>>> org.glassfish.grizzly.filterchain.DefaultFilterChain.process(DefaultFilt
>>>>>>> erChain.java:114) ~[grizzly-framework-2.3.5.jar:2.3.5]
>>>>>>> at
>>>>>>> org.glassfish.grizzly.ProcessorExecutor.execute(ProcessorExecutor.java:7
>>>>>>> 7) ~[grizzly-framework-2.3.5.jar:2.3.5]
>>>>>>> at
>>>>>>> org.glassfish.grizzly.filterchain.FilterChainContext.write(FilterChainCo
>>>>>>> ntext.java:853) ~[grizzly-framework-2.3.5.jar:2.3.5]
>>>>>>> at
>>>>>>> org.glassfish.grizzly.filterchain.FilterChainContext.write(FilterChainCo
>>>>>>> ntext.java:720) ~[grizzly-framework-2.3.5.jar:2.3.5]
>>>>>>> at
>>>>>>> com.ning.http.client.providers.grizzly.FeedableBodyGenerator.flushQueue(
>>>>>>> FeedableBodyGenerator.java:133)
>>>>>>> ~[async-http-client-1.7.20-SNAPSHOT.jar:na]
>>>>>>> at
>>>>>>> com.ning.http.client.providers.grizzly.FeedableBodyGenerator.feed(Feedab
>>>>>>> leBodyGenerator.java:95) ~[async-http-client-1.7.20-SNAPSHOT.jar:na]
>>>>>>>
>>>>>>> I think I got some other different stacks saying Connection Closed
>>>>>>> Remotely or something like that.
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>> So it seems that something is bound to my thread, and it stays bound to
>>>>>>> it for about 10 seconds, do you have any idea what it could be?
>>>>>>> (My connection timeout setting seems to have no effect on this 10s
>>>>>>> threshold)
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>> 2013/9/5 Ryan Lubke <ryan.lubke_at_oracle.com>
>>>>>>> That is one solution. I'm working out an alternative right now. Stay
>>>>>>> tuned!
>>>>>>>
>>>>>>>
>>>>>>> Anyway it's not a problem, I think the FeedableBodyGenerator.feed()
>>>>>>> method just has to block until a context has been (and ThreadCache
>>>>>>> initialized) to avoid OOM errors
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>> 2013/9/5 Sébastien Lorber <lorber.sebastien_at_gmail.com>
>>>>>>>
>>>>>>> What is very strange is that I tested with/without the same sessionCode
>>>>>>> with our previous code, the one not using FeedableBodyGenerator, which
>>>>>>> has a high memory consumption.
>>>>>>> Despites the fact it had high memory consumption, it seems work fine to
>>>>>>> upload multiple documents if allocated with a large heap, and the
>>>>>>> sessionCode seems to have no effect.
>>>>>>>
>>>>>>> On the new impl using the FeedableBodyGenerator, the sessionCode sent as
>>>>>>> a multipart bodypart seems to have an effect.
>>>>>>>
>>>>>>> I have tried to feed the queue before sending the request to AHC, but
>>>>>>> this leads to this exception (with/without sessionCode switching)
>>>>>>> Caused by: java.util.concurrent.TimeoutException: Timeout exceeded
>>>>>>> at
>>>>>>> com.ning.http.client.providers.grizzly.GrizzlyAsyncHttpProvider.timeout(
>>>>>>> GrizzlyAsyncHttpProvider.java:528)
>>>>>>> at
>>>>>>> com.ning.http.client.providers.grizzly.GrizzlyAsyncHttpProvider$3.onTime
>>>>>>> out(GrizzlyAsyncHttpProvider.java:361)
>>>>>>> at
>>>>>>> org.glassfish.grizzly.utils.IdleTimeoutFilter$DefaultWorker.doWork(IdleT
>>>>>>> imeoutFilter.java:383)
>>>>>>> at
>>>>>>> org.glassfish.grizzly.utils.IdleTimeoutFilter$DefaultWorker.doWork(IdleT
>>>>>>> imeoutFilter.java:362)
>>>>>>> at
>>>>>>> org.glassfish.grizzly.utils.DelayedExecutor$DelayedRunnable.run(DelayedE
>>>>>>> xecutor.java:158)
>>>>>>> at
>>>>>>> java.util.concurrent.ThreadPoolExecutor.runWorker(ThreadPoolExecutor.jav
>>>>>>> a:1110)
>>>>>>> at
>>>>>>> java.util.concurrent.ThreadPoolExecutor$Worker.run(ThreadPoolExecutor.ja
>>>>>>> va:603)
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>> 2013/9/5 Sébastien Lorber <lorber.sebastien_at_gmail.com>
>>>>>>> By the way, by using a low timeout with the same sessioncode, I got the
>>>>>>> following NPE:
>>>>>>>
>>>>>>> Caused by: java.lang.NullPointerException
>>>>>>> at
>>>>>>> com.ning.http.client.providers.grizzly.FeedableBodyGenerator.block(Feeda
>>>>>>> bleBodyGenerator.java:184)
>>>>>>> at
>>>>>>> com.ning.http.client.providers.grizzly.FeedableBodyGenerator.blockUntilQ
>>>>>>> ueueFree(FeedableBodyGenerator.java:167)
>>>>>>> at
>>>>>>> com.ning.http.client.providers.grizzly.FeedableBodyGenerator.flushQueue(
>>>>>>> FeedableBodyGenerator.java:124)
>>>>>>> at
>>>>>>> com.ning.http.client.providers.grizzly.FeedableBodyGenerator.feed(Feedab
>>>>>>> leBodyGenerator.java:94)
>>>>>>>
>>>>>>> GrizzlyAsyncHttpProvider.HttpTransactionContext httpCtx =
>>>>>>> getHttpTransactionContext(c);
>>>>>>> httpCtx.abort(e.getCause());
>>>>>>>
>>>>>>> I guess the httpCtx is not already available to be aborted
>>>>>>>
>>>>>>>
>>>>>>> 2013/9/5 Sébastien Lorber <lorber.sebastien_at_gmail.com>
>>>>>>>
>>>>>>>
>>>>>>> This is right, here's a log I have when I use the same session code, ie
>>>>>>> the remote host is blocking the data or something.
>>>>>>> This is obtained by running 5 parallel uploads.
>>>>>>>
>>>>>>> Flushing queue of size 0 with allowBlocking = false
>>>>>>> Flushing queue of size 1 with allowBlocking = true
>>>>>>> Flushing queue of size 97 with allowBlocking = false
>>>>>>> Flushing queue of size 100 with allowBlocking = false
>>>>>>> Flushing queue of size 160 with allowBlocking = true
>>>>>>> Flushing queue of size 1 with allowBlocking = true
>>>>>>> Flushing queue of size 1 with allowBlocking = true
>>>>>>> Flushing queue of size 1 with allowBlocking = true
>>>>>>> Flushing queue of size 1 with allowBlocking = true
>>>>>>> Flushing queue of size 1 with allowBlocking = true
>>>>>>> Flushing queue of size 0 with allowBlocking = true
>>>>>>> Flushing queue of size 1 with allowBlocking = true
>>>>>>> Flushing queue of size 1 with allowBlocking = true
>>>>>>> Flushing queue of size 1 with allowBlocking = true
>>>>>>> Flushing queue of size 1 with allowBlocking = true
>>>>>>> Flushing queue of size 1 with allowBlocking = true
>>>>>>> Flushing queue of size 1 with allowBlocking = true
>>>>>>> Flushing queue of size 1 with allowBlocking = true
>>>>>>> Flushing queue of size 1 with allowBlocking = true
>>>>>>> Flushing queue of size 1 with allowBlocking = true
>>>>>>> Flushing queue of size 1 with allowBlocking = true
>>>>>>> Flushing queue of size 1 with allowBlocking = true
>>>>>>> Flushing queue of size 1 with allowBlocking = true
>>>>>>> Flushing queue of size 1 with allowBlocking = true
>>>>>>> Flushing queue of size 1 with allowBlocking = true
>>>>>>> Flushing queue of size 1 with allowBlocking = true
>>>>>>> Flushing queue of size 1 with allowBlocking = true
>>>>>>> Flushing queue of size 1 with allowBlocking = true
>>>>>>> Flushing queue of size 1 with allowBlocking = true
>>>>>>> Flushing queue of size 1 with allowBlocking = true
>>>>>>> Flushing queue of size 1 with allowBlocking = true
>>>>>>> Flushing queue of size 1 with allowBlocking = true
>>>>>>> Flushing queue of size 1 with allowBlocking = true
>>>>>>> Flushing queue of size 1 with allowBlocking = true
>>>>>>> Flushing queue of size 1 with allowBlocking = true
>>>>>>> Flushing queue of size 1 with allowBlocking = true
>>>>>>> Flushing queue of size 1 with allowBlocking = true
>>>>>>> Flushing queue of size 1 with allowBlocking = true
>>>>>>> Flushing queue of size 1 with allowBlocking = true
>>>>>>> Flushing queue of size 1 with allowBlocking = true
>>>>>>> Flushing queue of size 1 with allowBlocking = true
>>>>>>> Flushing queue of size 1 with allowBlocking = true
>>>>>>> Flushing queue of size 1 with allowBlocking = true
>>>>>>> Flushing queue of size 1 with allowBlocking = true
>>>>>>> Flushing queue of size 1 with allowBlocking = true
>>>>>>> Flushing queue of size 1 with allowBlocking = true
>>>>>>> Flushing queue of size 1 with allowBlocking = true
>>>>>>> Flushing queue of size 1 with allowBlocking = true
>>>>>>> Flushing queue of size 1 with allowBlocking = true
>>>>>>> Flushing queue of size 1 with allowBlocking = true
>>>>>>> Flushing queue of size 1 with allowBlocking = true
>>>>>>> Flushing queue of size 1 with allowBlocking = true
>>>>>>> Flushing queue of size 1 with allowBlocking = true
>>>>>>> Flushing queue of size 1 with allowBlocking = true
>>>>>>> Flushing queue of size 1 with allowBlocking = true
>>>>>>> Flushing queue of size 1 with allowBlocking = true
>>>>>>> java.lang.OutOfMemoryError: GC overhead limit exceeded
>>>>>>> Dumping heap to /ome/lorber/ureau/om ...
>>>>>>> Unable to create /ome/lorber/ureau/om: Le fichier existe
>>>>>>> Flushing queue of size 1 with allowBlocking = true
>>>>>>> Flushing queue of size 1 with allowBlocking = true
>>>>>>> Flushing queue of size 1 with allowBlocking = true
>>>>>>> Flushing queue of size 1 with allowBlocking = true
>>>>>>> Flushing queue of size 1 with allowBlocking = true
>>>>>>> Disconnected from the target VM, address: '127.0.0.1:49268
>>>>>>> <http://127.0.0.1:49268> ', transport: 'socket'
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>> Otherwise, with different session codes, I get the following:
>>>>>>>
>>>>>>> Flushing queue of size 0 with allowBlocking = false
>>>>>>> Flushing queue of size 1 with allowBlocking = true
>>>>>>> Flushing queue of size 1 with allowBlocking = true
>>>>>>> Flushing queue of size 1 with allowBlocking = true
>>>>>>> Flushing queue of size 0 with allowBlocking = false
>>>>>>> Flushing queue of size 1 with allowBlocking = true
>>>>>>> Flushing queue of size 1 with allowBlocking = true
>>>>>>> Flushing queue of size 0 with allowBlocking = false
>>>>>>> Flushing queue of size 0 with allowBlocking = false
>>>>>>> Flushing queue of size 1 with allowBlocking = true
>>>>>>> Flushing queue of size 1 with allowBlocking = true
>>>>>>> Flushing queue of size 1 with allowBlocking = true
>>>>>>> Flushing queue of size 1 with allowBlocking = true
>>>>>>> Flushing queue of size 1 with allowBlocking = true
>>>>>>> Flushing queue of size 1 with allowBlocking = true
>>>>>>> Flushing queue of size 1 with allowBlocking = true
>>>>>>> Flushing queue of size 1 with allowBlocking = true
>>>>>>> Flushing queue of size 1 with allowBlocking = true
>>>>>>> ... and this continues without OOM
>>>>>>>
>>>>>>>
>>>>>>> So, this seems to be the problem.
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>> I think it would be great to be able to be able to choose the queue impl
>>>>>>> behind that FeedableBodyGenerator, like I suggested in my pull request.
>>>>>>>
>>>>>>> See here:
>>>>>>> https://github.com/slorber/async-http-client/blob/79b0c3b28a61b0aa4c4b05
>>>>>>> 5bca8f6be11d9ed1e6/src/main/java/com/ning/http/client/providers/grizzly/
>>>>>>> FeedableBodyGenerator.java
>>>>>>>
>>>>>>> Using a LinkedBlockingQueue seems to be a nice idea in this context, and
>>>>>>> in my case I would probably use a queue of size 1
>>>>>>>
>>>>>>>
>>>>>>> This would handle the blocking of the feed method, without having to
>>>>>>> use this:
>>>>>>> if (context != null) {
>>>>>>> flushQueue(true);
>>>>>>> }
>>>>>>>
>>>>>>>
>>>>>>> Or perhaps the feed() method have to wait until a context is set in the
>>>>>>> BodyGenerator ?
>>>>>>>
>>>>>>>
>>>>>>> I think it would be more clear if the initializeAsynchronousTransfer
>>>>>>> simply didn't flush the queue but just setup the context.
>>>>>>> Then the feed method would block until there's a context set, and then
>>>>>>> flush the queue with blocking behavior.
>>>>>>>
>>>>>>> This is probably the next step, but as we are using AHC for async, it
>>>>>>> would probably be great if that blocking feed() method was called in a
>>>>>>> worker thread instead of our main thread.
>>>>>>> I won't use this but someone who really wants a non-blocking impl of
>>>>>>> performant multipart fileupload would probably need this, or will use an
>>>>>>> ExecutorService for the feeding operations as a workaround.
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>> Thanks again for your reactivity
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>> 2013/9/4 Ryan Lubke <ryan.lubke_at_oracle.com>
>>>>>>>
>>>>>>>
>>>>>>> Sébastien Lorber wrote:
>>>>>>>
>>>>>>> I've integrated this change and it works fine except a little detail.
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>> I'm uploading files to a third party API (a bit like S3).
>>>>>>> This API requires a "sessionCode" in each request. So there is a
>>>>>>> multipart StringPart with that SessionCode.
>>>>>>>
>>>>>>> We used to have a cache which holds the sessionCode 30min per user so
>>>>>>> that we do not need to init a new session each time.
>>>>>>>
>>>>>>> I had troubles in this very specific case: when I upload 5 docs with the
>>>>>>> same session code.
>>>>>>> When I remove the cache and use 5 different session codes, it works
>>>>>>> fine.
>>>>>>>
>>>>>>> I guess the remote service is blocking concurrent uploads with the same
>>>>>>> session code. I don't know at all.
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>> Where I want to go is that I wouldn't have expected Grizzly to OOM
>>>>>>>
>>>>>>>
>>>>>>> Avertissement: Exception during FilterChain execution
>>>>>>> java.lang.OutOfMemoryError: Java heap space
>>>>>>> at java.nio.HeapByteBuffer.<init>(HeapByteBuffer.java:57)
>>>>>>> at java.nio.ByteBuffer.allocate(ByteBuffer.java:331)
>>>>>>> at
>>>>>>> org.glassfish.grizzly.ssl.SSLUtils.allocateOutputBuffer(SSLUtils.java:34
>>>>>>> 2)
>>>>>>> at
>>>>>>> org.glassfish.grizzly.ssl.SSLBaseFilter$2.grow(SSLBaseFilter.java:117)
>>>>>>> at
>>>>>>> org.glassfish.grizzly.ssl.SSLConnectionContext.ensureBufferSize(SSLConne
>>>>>>> ctionContext.java:392)
>>>>>>> at
>>>>>>> org.glassfish.grizzly.ssl.SSLConnectionContext.wrap(SSLConnectionContext
>>>>>>> .java:272)
>>>>>>> at
>>>>>>> org.glassfish.grizzly.ssl.SSLConnectionContext.wrapAll(SSLConnectionCont
>>>>>>> ext.java:238)
>>>>>>> at
>>>>>>> org.glassfish.grizzly.ssl.SSLBaseFilter.wrapAll(SSLBaseFilter.java:405)
>>>>>>> at
>>>>>>> org.glassfish.grizzly.ssl.SSLBaseFilter.handleWrite(SSLBaseFilter.java:3
>>>>>>> 20)
>>>>>>> at org.glassfish.grizzly.ssl.SSLFilter.accurateWrite(SSLFilter.java:263)
>>>>>>> at org.glassfish.grizzly.ssl.SSLFilter.handleWrite(SSLFilter.java:143)
>>>>>>> at
>>>>>>> com.ning.http.client.providers.grizzly.GrizzlyAsyncHttpProvider$Switchin
>>>>>>> gSSLFilter.handleWrite(GrizzlyAsyncHttpProvider.java:2500)
>>>>>>> at
>>>>>>> org.glassfish.grizzly.filterchain.ExecutorResolver$8.execute(ExecutorRes
>>>>>>> olver.java:111)
>>>>>>> at
>>>>>>> org.glassfish.grizzly.filterchain.DefaultFilterChain.executeFilter(Defau
>>>>>>> ltFilterChain.java:288)
>>>>>>> at
>>>>>>> org.glassfish.grizzly.filterchain.DefaultFilterChain.executeChainPart(De
>>>>>>> faultFilterChain.java:206)
>>>>>>> at
>>>>>>> org.glassfish.grizzly.filterchain.DefaultFilterChain.execute(DefaultFilt
>>>>>>> erChain.java:136)
>>>>>>> at
>>>>>>> org.glassfish.grizzly.filterchain.DefaultFilterChain.process(DefaultFilt
>>>>>>> erChain.java:114)
>>>>>>> at
>>>>>>> org.glassfish.grizzly.ProcessorExecutor.execute(ProcessorExecutor.java:7
>>>>>>> 7)
>>>>>>> at
>>>>>>> org.glassfish.grizzly.filterchain.FilterChainContext$1.run(FilterChainCo
>>>>>>> ntext.java:196)
>>>>>>> at
>>>>>>> org.glassfish.grizzly.filterchain.FilterChainContext.resume(FilterChainC
>>>>>>> ontext.java:220)
>>>>>>> at
>>>>>>> org.glassfish.grizzly.ssl.SSLFilter$SSLHandshakeContext.completed(SSLFil
>>>>>>> ter.java:383)
>>>>>>> at
>>>>>>> org.glassfish.grizzly.ssl.SSLFilter.notifyHandshakeComplete(SSLFilter.ja
>>>>>>> va:278)
>>>>>>> at
>>>>>>> org.glassfish.grizzly.ssl.SSLBaseFilter.handleRead(SSLBaseFilter.java:27
>>>>>>> 5)
>>>>>>> at
>>>>>>> com.ning.http.client.providers.grizzly.GrizzlyAsyncHttpProvider$Switchin
>>>>>>> gSSLFilter.handleRead(GrizzlyAsyncHttpProvider.java:2490)
>>>>>>> at
>>>>>>> org.glassfish.grizzly.filterchain.ExecutorResolver$9.execute(ExecutorRes
>>>>>>> olver.java:119)
>>>>>>> at
>>>>>>> org.glassfish.grizzly.filterchain.DefaultFilterChain.executeFilter(Defau
>>>>>>> ltFilterChain.java:288)
>>>>>>> at
>>>>>>> org.glassfish.grizzly.filterchain.DefaultFilterChain.executeChainPart(De
>>>>>>> faultFilterChain.java:206)
>>>>>>> at
>>>>>>> org.glassfish.grizzly.filterchain.DefaultFilterChain.execute(DefaultFilt
>>>>>>> erChain.java:136)
>>>>>>> at
>>>>>>> org.glassfish.grizzly.filterchain.DefaultFilterChain.process(DefaultFilt
>>>>>>> erChain.java:114)
>>>>>>> at
>>>>>>> org.glassfish.grizzly.ProcessorExecutor.execute(ProcessorExecutor.java:7
>>>>>>> 7)
>>>>>>> at
>>>>>>> org.glassfish.grizzly.nio.transport.TCPNIOTransport.fireIOEvent(TCPNIOTr
>>>>>>> ansport.java:546)
>>>>>>> at
>>>>>>> org.glassfish.grizzly.strategies.AbstractIOStrategy.fireIOEvent(Abstract
>>>>>>> IOStrategy.java:113)
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>> Caused by: java.util.concurrent.TimeoutException: null
>>>>>>> at
>>>>>>> org.glassfish.grizzly.impl.SafeFutureImpl$Sync.innerGet(SafeFutureImpl.j
>>>>>>> ava:367) ~[grizzly-framework-2.3.5.jar:2.3.5]
>>>>>>> at
>>>>>>> org.glassfish.grizzly.impl.SafeFutureImpl.get(SafeFutureImpl.java:274)
>>>>>>> ~[grizzly-framework-2.3.5.jar:2.3.5]
>>>>>>> at
>>>>>>> com.ning.http.client.providers.grizzly.FeedableBodyGenerator.block(Feeda
>>>>>>> bleBodyGenerator.java:177) ~[async-http-client-1.7.20-204092c.jar:na]
>>>>>>> at
>>>>>>> com.ning.http.client.providers.grizzly.FeedableBodyGenerator.blockUntilQ
>>>>>>> ueueFree(FeedableBodyGenerator.java:167)
>>>>>>> ~[async-http-client-1.7.20-204092c.jar:na]
>>>>>>> at
>>>>>>> com.ning.http.client.providers.grizzly.FeedableBodyGenerator.flushQueue(
>>>>>>> FeedableBodyGenerator.java:124)
>>>>>>> ~[async-http-client-1.7.20-204092c.jar:na]
>>>>>>> at
>>>>>>> com.ning.http.client.providers.grizzly.FeedableBodyGenerator.feed(Feedab
>>>>>>> leBodyGenerator.java:94) ~[async-http-client-1.7.20-204092c.jar:na]
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>> multipart.body.generator.feeder.buffer=100000 -> size of each Buffer
>>>>>>> sent to the FeedableBodyGenerator
>>>>>>> transport.max.pending.bytes=1000000
>>>>>>> (I tried other settings, including AUTO_SIZE)
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>> Do you have any idea why is there an OOM with these settings?
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>> Perhaps it is because the feed() method of FeedableBodyGenerator doesn't
>>>>>>> block until the context is initialized.
>>>>>>> I guess the initializeAsynchronousTransfer is called only once the
>>>>>>> connection is established, and perhaps a lot of Buffer are added to the
>>>>>>> queue...
>>>>>>> Yes, it's invoked once the request has been dispatched, so if the
>>>>>>> generator is fed a lot before the request, I could see this happening.
>>>>>>> I'll see what I can do to alleviate that case.
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>> But I'm not sure at all because the session code is transmitted as a
>>>>>>> BodyPart and I get the same problem if i put it as the first or last
>>>>>>> multipart.
>>>>>>>
>>>>>>> It's not a big deal, perhaps I should always use a different session
>>>>>>> code for concurrent operations but I'd like to be sure that we won't
>>>>>>> have this issue in production...
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>> 2013/9/3 Ryan Lubke <ryan.lubke_at_oracle.com>
>>>>>>> Good catch. Fixed.
>>>>>>>
>>>>>>>
>>>>>>> Sébastien Lorber wrote:
>>>>>>>
>>>>>>>
>>>>>>> Hello,
>>>>>>>
>>>>>>>
>>>>>>> There's a little mistake in the grizzly ahc provider relative to the
>>>>>>> write queue size.
>>>>>>>
>>>>>>>
>>>>>>> https://github.com/AsyncHttpClient/async-http-client/blob/b5d97efe9fe141
>>>>>>> 13ea92fb1f7db192a2d090fad7/src/main/java/com/ning/http/client/providers/
>>>>>>> grizzly/GrizzlyAsyncHttpProvider.java#L419
>>>>>>>
>>>>>>> As you can see, the TransportCustomizer is called, and then the write
>>>>>>> queue size (among other things) is set to AUTO_SIZE (instead of
>>>>>>> previously UNLIMITED)
>>>>>>>
>>>>>>> clientTransport.getAsyncQueueIO().getWriter()
>>>>>>>
>>>>>>> .setMaxPendingBytesPerConnection(AsyncQueueWriter.AUTO_SIZE);
>>>>>>>
>>>>>>>
>>>>>>> I think the default settings like this AUTO_SIZE attribute should be set
>>>>>>> before the customization of the transport, or they would override the
>>>>>>> value we customized.
>>>>>>>
>>>>>>>
>>>>>>> This is actually my case, since I can't reproduce my "bug" which is
>>>>>>> "high memory consumption", even when using -1 / UNLIMITED in the
>>>>>>> TransportCustomizer.
>>>>>>>
>>>>>>>
>>>>>>> This could work fine for me with AUTO_SIZE, but I'd rather be able to
>>>>>>> tune this parameter during load tests to see the effect.
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>> 2013/8/31 Sebastien Lorber <lorber.sebastien_at_gmail.com>
>>>>>>> Thanks i will ckeck that on monday. I can now upload a 500m file with
>>>>>>> 40m heap size ;)
>>>>>>>
>>>>>>> Envoyé de mon iPhone
>>>>>>>
>>>>>>> Le 30 août 2013 à 20:49, Ryan Lubke <ryan.lubke_at_oracle.com> a écrit :
>>>>>>>
>>>>>>> I'm going to be updating the Grizzly provider such that AUTO_SIZE (not
>>>>>>> AUTO_TUNE) is the default, so you can avoid the use of the
>>>>>>> TransportCustomizer.
>>>>>>>
>>>>>>> Ryan Lubke wrote:
>>>>>>> Regarding your tuning question, I would probably set the value to
>>>>>>> AsyncQueueWriter.AUTO_TUNE (this will be four times the socket write
>>>>>>> buffer) and see how that works.
>>>>>>>
>>>>>>> Ryan Lubke wrote:
>>>>>>> A question first. With these changes, your memory usage is more inline
>>>>>>> with what you were looking for?
>>>>>>>
>>>>>>> Sébastien Lorber wrote:
>>>>>>> By the way, do you have any idea when the 1.7.20 will be released (with
>>>>>>> these new improvements?)
>>>>>>>
>>>>>>> We would like to know if we wait for a release or if we install our own
>>>>>>> temp release on Nexus :)
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>> 2013/8/30 Sébastien Lorber <lorber.sebastien_at_gmail.com>
>>>>>>> Thank you, it works fine!
>>>>>>>
>>>>>>>
>>>>>>> I just had to modify a single line after your commit.
>>>>>>>
>>>>>>> com.ning.http.client.providers.grizzly.GrizzlyAsyncHttpProvider#initiali
>>>>>>> zeTransport
>>>>>>> ->
>>>>>>> clientTransport.getAsyncQueueIO().getWriter().setMaxPendingBytesPerConne
>>>>>>> ction(10000);
>>>>>>>
>>>>>>>
>>>>>>> If I let the initial value (-1) it won't block, canWrite always returns
>>>>>>> true
>>>>>>>
>>>>>>>
>>>>>>> Btw, on AHC I didn't find any way to pass this value as a config
>>>>>>> attribute, neither the size of the write buffer you talked about.
>>>>>>>
>>>>>>> So in the end, is there a way with current AHC code to use this
>>>>>>> "canWrite = false" behavior?
>>>>>>> If not, can you please provide a way to set this behavior on v1.7.20 ?
>>>>>>> thanks
>>>>>>>
>>>>>>>
>>>>>>> PS: does it make sens to use the same number of bytes un the
>>>>>>> feed(Buffer) method and in the setMaxPendingBytesPerConnection(10000); ?
>>>>>>> do you have any tuning recommandation?
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>> 2013/8/29 Ryan Lubke <ryan.lubke_at_oracle.com>
>>>>>>> Please disregard.
>>>>>>>
>>>>>>>
>>>>>>> Ryan Lubke wrote:
>>>>>>> Sébastien,
>>>>>>>
>>>>>>> Could you also provide a sample of how you're performing your feed?
>>>>>>>
>>>>>>> Thanks,
>>>>>>> -rl
>>>>>>>
>>>>>>> Ryan Lubke wrote:
>>>>>>> Sébastien,
>>>>>>>
>>>>>>> I'd recommend looking at Connection.canWrite() [1] and
>>>>>>> Connection.notifyCanWrite(WriteListener) [1]
>>>>>>>
>>>>>>> By default, Grizzly will configure the async write queue length to be
>>>>>>> four times the write buffer size (which is based off the socket write
>>>>>>> buffer).
>>>>>>> When this queue exceeds this value, canWrite() will return false.
>>>>>>>
>>>>>>> When this occurs, you can register a WriteListener to be notified when
>>>>>>> the queue length is below the configured max and then simulate blocking
>>>>>>> until the onWritePossible() callback has been invoked.
>>>>>>>
>>>>>>> ----------------------------------------------------------------
>>>>>>>
>>>>>>> final FutureImpl<Boolean> future = Futures.createSafeFuture();
>>>>>>>
>>>>>>> // Connection may be obtained by calling
>>>>>>> FilterChainContext.getConnection().
>>>>>>> connection.notifyCanWrite(new WriteHandler() {
>>>>>>>
>>>>>>> @Override
>>>>>>> public void onWritePossible() throws Exception {
>>>>>>> future.result(Boolean.TRUE);
>>>>>>> }
>>>>>>>
>>>>>>> @Override
>>>>>>> public void onError(Throwable t) {
>>>>>>> future.failure(Exceptions.makeIOException(t));
>>>>>>> }
>>>>>>> });
>>>>>>>
>>>>>>> try {
>>>>>>> final long writeTimeout = 30;
>>>>>>> future.get(writeTimeout, TimeUnit.MILLISECONDS);
>>>>>>> } catch (ExecutionException e) {
>>>>>>> HttpTransactionContext httpCtx =
>>>>>>> HttpTransactionContext.get(connection);
>>>>>>> httpCtx.abort(e.getCause());
>>>>>>> } catch (Exception e) {
>>>>>>> HttpTransactionContext httpCtx =
>>>>>>> HttpTransactionContext.get(connection);
>>>>>>> httpCtx.abort(e);
>>>>>>> }
>>>>>>>
>>>>>>> ---------------------------------------------------------------------
>>>>>>>
>>>>>>> [1]
>>>>>>> http://grizzly.java.net/docs/2.3/apidocs/org/glassfish/grizzly/OutputSin
>>>>>>> k.html.
>>>>>>>
>>>>>>> Sébastien Lorber wrote:
>>>>>>> Ryan, I've did some other tests.
>>>>>>>
>>>>>>>
>>>>>>> It seems that using a blocking queue in the FeedableBodyGenerator is
>>>>>>> totally useless because the thread consuming it is not blocking and the
>>>>>>> queue never blocks the feeding, which was my intention in the beginning.
>>>>>>> Maybe it depends on the IO strategy used?
>>>>>>> I use AHC default which seems to use SameThreadIOStrategy so I don't
>>>>>>> think it's related to the IO strategy.
>>>>>>>
>>>>>>>
>>>>>>> So in the end I can upload a 70m file with a heap of 50m, but I have to
>>>>>>> put a Thread.sleep(30) between each 100k Buffer send to the
>>>>>>> FeedableBodyGenerator
>>>>>>>
>>>>>>> The connection with the server is not good here, but in production it is
>>>>>>> normally a lot better as far as I know.
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>> I've tried things like
>>>>>>> clientTransport.getAsyncQueueIO().getWriter().setMaxPendingBytesPerConne
>>>>>>> ction(100000); but it doesn't seem to work for me.
>>>>>>>
>>>>>>> I'd like the Grizzly internals to block when there are too much pending
>>>>>>> bytes to send. Is it possible?
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>> PS: I've just been able to send a 500mo file with 100mo heap, but it
>>>>>>> needed a sleep of 100ms between each 100k Buffer sent to the
>>>>>>> bodygenerator
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>> 2013/8/29 Sébastien Lorber <lorber.sebastien_at_gmail.com>
>>>>>>>
>>>>>>>
>>>>>>> By chance do you if I can remove the MessageCloner used in the SSL
>>>>>>> filter?
>>>>>>> SSLBaseFilter$OnWriteCopyCloner
>>>>>>>
>>>>>>> It seems to allocate a lot of memory.
>>>>>>> I don't really understand why messages have to be cloned, can I remove
>>>>>>> this? How?
>>>>>>>
>>>>>>>
>>>>>>> 2013/8/29 Sébastien Lorber <lorber.sebastien_at_gmail.com>
>>>>>>>
>>>>>>> I'm trying to send a 500m file for my tests with a heap of 400m.
>>>>>>>
>>>>>>> In our real use cases we would probably have files under 20mo but we
>>>>>>> want to reduce the memory consumption because we can have x parallel
>>>>>>> uploads on the same server according to the user activity.
>>>>>>>
>>>>>>> I'll try to check if using this BodyGenerator reduced the memory
>>>>>>> footprint or if it's almost like before.
>>>>>>>
>>>>>>>
>>>>>>> 2013/8/28 Ryan Lubke <ryan.lubke_at_oracle.com>
>>>>>>> At this point in time, as far as the SSL buffer allocation is concerned,
>>>>>>> it's untunable.
>>>>>>>
>>>>>>> That said, feel free to open a feature request.
>>>>>>>
>>>>>>> As to your second question, there is no suggested size. This is all
>>>>>>> very application specific.
>>>>>>>
>>>>>>> I'm curious, how large of a file are you sending?
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>> Sébastien Lorber wrote:
>>>>>>> I have seen a lot of buffers which have a size of 33842 and it seems the
>>>>>>> limit is near half the capacity.
>>>>>>>
>>>>>>> Perhaps there's a way to tune that buffer size so that it consumes less
>>>>>>> memory?
>>>>>>> Is there an ideal Buffer size to send to the feed method?
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>> 2013/8/28 Ryan Lubke <ryan.lubke_at_oracle.com>
>>>>>>> I'll be reviewing the PR today, thanks again!
>>>>>>>
>>>>>>> Regarding the OOM: as it stands now, for each new buffer that is passed
>>>>>>> to the SSLFilter, we allocate a buffer twice the size in order to
>>>>>>> accommodate the encrypted result. So there's an increase.
>>>>>>>
>>>>>>> Depending on the socket configurations of both endpoints, and how fast
>>>>>>> the remote is reading data, it could
>>>>>>> be the write queue is becoming too large. We do have a way to detect
>>>>>>> this situation, but I'm pretty sure
>>>>>>> the Grizzly internals are currently shielded here. I will see what I
>>>>>>> can do to allow users to leverage this.
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>> Sébastien Lorber wrote:
>>>>>>> Hello,
>>>>>>>
>>>>>>> I've made my pull request.
>>>>>>> https://github.com/AsyncHttpClient/async-http-client/pull/367
>>>>>>>
>>>>>>> With my usecase it works, the file is uploaded like before.
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>> But I didn't notice a big memory improvement.
>>>>>>>
>>>>>>> Is it possible that SSL doesn't allow to stream the body or something
>>>>>>> like that?
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>> In memory, I have a lot of:
>>>>>>> - HeapByteBuffer
>>>>>>> Which are hold by SSLUtils$3
>>>>>>> Which are hold by BufferBuffers
>>>>>>> Which are hold by WriteResult
>>>>>>> Which are hold by AsyncWriteQueueRecord
>>>>>>>
>>>>>>>
>>>>>>> Here is an exemple of the OOM stacktrace:
>>>>>>>
>>>>>>> java.lang.OutOfMemoryError: Java heap space
>>>>>>> at java.nio.HeapByteBuffer.<init>(HeapByteBuffer.java:57)
>>>>>>> at java.nio.ByteBuffer.allocate(ByteBuffer.java:331)
>>>>>>> at
>>>>>>> org.glassfish.grizzly.ssl.SSLUtils.allocateOutputBuffer(SSLUtils.java:34
>>>>>>> 2)
>>>>>>> at
>>>>>>> org.glassfish.grizzly.ssl.SSLBaseFilter$2.grow(SSLBaseFilter.java:117)
>>>>>>> at
>>>>>>> org.glassfish.grizzly.ssl.SSLConnectionContext.ensureBufferSize(SSLConne
>>>>>>> ctionContext.java:392)
>>>>>>> at
>>>>>>> org.glassfish.grizzly.ssl.SSLConnectionContext.wrap(SSLConnectionContext
>>>>>>> .java:272)
>>>>>>> at
>>>>>>> org.glassfish.grizzly.ssl.SSLConnectionContext.wrapAll(SSLConnectionCont
>>>>>>> ext.java:227)
>>>>>>> at
>>>>>>> org.glassfish.grizzly.ssl.SSLBaseFilter.wrapAll(SSLBaseFilter.java:404)
>>>>>>> at
>>>>>>> org.glassfish.grizzly.ssl.SSLBaseFilter.handleWrite(SSLBaseFilter.java:3
>>>>>>> 19)
>>>>>>> at org.glassfish.grizzly.ssl.SSLFilter.accurateWrite(SSLFilter.java:255)
>>>>>>> at org.glassfish.grizzly.ssl.SSLFilter.handleWrite(SSLFilter.java:143)
>>>>>>> at
>>>>>>> com.ning.http.client.providers.grizzly.GrizzlyAsyncHttpProvider$Switchin
>>>>>>> gSSLFilter.handleWrite(GrizzlyAsyncHttpProvider.java:2503)
>>>>>>> at
>>>>>>> org.glassfish.grizzly.filterchain.ExecutorResolver$8.execute(ExecutorRes
>>>>>>> olver.java:111)
>>>>>>> at
>>>>>>> org.glassfish.grizzly.filterchain.DefaultFilterChain.executeFilter(Defau
>>>>>>> ltFilterChain.java:288)
>>>>>>> at
>>>>>>> org.glassfish.grizzly.filterchain.DefaultFilterChain.executeChainPart(De
>>>>>>> faultFilterChain.java:206)
>>>>>>> at
>>>>>>> org.glassfish.grizzly.filterchain.DefaultFilterChain.execute(DefaultFilt
>>>>>>> erChain.java:136)
>>>>>>> at
>>>>>>> org.glassfish.grizzly.filterchain.DefaultFilterChain.process(DefaultFilt
>>>>>>> erChain.java:114)
>>>>>>> at
>>>>>>> org.glassfish.grizzly.ProcessorExecutor.execute(ProcessorExecutor.java:7
>>>>>>> 7)
>>>>>>> at
>>>>>>> org.glassfish.grizzly.filterchain.FilterChainContext.write(FilterChainCo
>>>>>>> ntext.java:853)
>>>>>>> at
>>>>>>> org.glassfish.grizzly.filterchain.FilterChainContext.write(FilterChainCo
>>>>>>> ntext.java:720)
>>>>>>> at
>>>>>>> com.ning.http.client.providers.grizzly.FeedableBodyGenerator.flushQueue(
>>>>>>> FeedableBodyGenerator.java:132)
>>>>>>> at
>>>>>>> com.ning.http.client.providers.grizzly.FeedableBodyGenerator.feed(Feedab
>>>>>>> leBodyGenerator.java:101)
>>>>>>> at
>>>>>>> com.ning.http.client.providers.grizzly.MultipartBodyGeneratorFeeder$Feed
>>>>>>> BodyGeneratorOutputStream.write(MultipartBodyGeneratorFeeder.java:222)
>>>>>>> at
>>>>>>> java.io.BufferedOutputStream.flushBuffer(BufferedOutputStream.java:82)
>>>>>>> at java.io.BufferedOutputStream.write(BufferedOutputStream.java:126)
>>>>>>> at com.ning.http.multipart.FilePart.sendData(FilePart.java:179)
>>>>>>> at com.ning.http.multipart.Part.send(Part.java:331)
>>>>>>> at com.ning.http.multipart.Part.sendParts(Part.java:397)
>>>>>>> at
>>>>>>> com.ning.http.client.providers.grizzly.MultipartBodyGeneratorFeeder.feed
>>>>>>> (MultipartBodyGeneratorFeeder.java:144)
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>> Any idea?
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>> 2013/8/27 Ryan Lubke <ryan.lubke_at_oracle.com>
>>>>>>> Excellent! Looking forward to the pull request!
>>>>>>>
>>>>>>>
>>>>>>> Sébastien Lorber wrote:
>>>>>>> Ryan thanks, it works fine, I'll make a pull request on AHC tomorrow
>>>>>>> with a better code using the same Part classes that already exist.
>>>>>>>
>>>>>>> I created an OutputStream that redirects to the BodyGenerator feeder.
>>>>>>>
>>>>>>> The problem I currently have is that the feeder feeds the queue faster
>>>>>>> than the async thread polling it :)
>>>>>>> I need to expose a limit to that queue size or something, will work on
>>>>>>> that, it will be better than a thread sleep to slow down the filepart
>>>>>>> reading
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>> 2013/8/27 Ryan Lubke <ryan.lubke_at_oracle.com>
>>>>>>> Yes, something like that. I was going to tackle adding something like
>>>>>>> this today. I'll follow up with something you can test out.
>>>>>>>
>>>>>>>
>>>>>>> Sébastien Lorber wrote:
>>>>>>> Ok thanks!
>>>>>>>
>>>>>>> I think I see what I could do, probably something like that:
>>>>>>>
>>>>>>>
>>>>>>> FeedableBodyGenerator bodyGenerator = new FeedableBodyGenerator();
>>>>>>> MultipartBodyGeneratorFeeder bodyGeneratorFeeder = new
>>>>>>> MultipartBodyGeneratorFeeder(bodyGenerator);
>>>>>>> Request uploadRequest1 = new RequestBuilder("POST")
>>>>>>> .setUrl("url")
>>>>>>> .setBody(bodyGenerator)
>>>>>>> .build();
>>>>>>>
>>>>>>> ListenableFuture<Response> asyncRes = asyncHttpClient
>>>>>>> .prepareRequest(uploadRequest1)
>>>>>>> .execute(new AsyncCompletionHandlerBase());
>>>>>>>
>>>>>>>
>>>>>>> bodyGeneratorFeeder.append("param1","value1");
>>>>>>> bodyGeneratorFeeder.append("param2","value2");
>>>>>>> bodyGeneratorFeeder.append("fileToUpload",fileInputStream);
>>>>>>> bodyGeneratorFeeder.end();
>>>>>>>
>>>>>>> Response uploadResponse = asyncRes.get();
>>>>>>>
>>>>>>>
>>>>>>> Does it seem ok to you?
>>>>>>>
>>>>>>> I guess it could be interesting to provide that
>>>>>>> MultipartBodyGeneratorFeeder class to AHC or Grizzly since some other
>>>>>>> people may want to achieve the same thing
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>> 2013/8/26 Ryan Lubke <ryan.lubke_at_oracle.com>
>>>>>>>
>>>>>>>
>>>>>>> Sébastien Lorber wrote:
>>>>>>> Hello,
>>>>>>>
>>>>>>> I would like to know if it's possible to upload a file with AHC /
>>>>>>> Grizzly in streaming, I mean without loading the whole file bytes in
>>>>>>> memory.
>>>>>>>
>>>>>>> The default behavior seems to allocate a byte[] which contans the whole
>>>>>>> file, so it means that my server can be OOM if too many users upload a
>>>>>>> large file in the same time.
>>>>>>>
>>>>>>>
>>>>>>> I've tryied with a Heap and ByteBuffer memory managers, with
>>>>>>> reallocate=true/false but no more success.
>>>>>>>
>>>>>>> It seems the whole file content is appended wto the BufferOutputStream,
>>>>>>> and then the underlying buffer is written.
>>>>>>>
>>>>>>> At least this seems to be the case with AHC integration:
>>>>>>> https://github.com/AsyncHttpClient/async-http-client/blob/6faf1f316e5546
>>>>>>> 110b0779a5a42fd9d03ba6bc15/providers/grizzly/src/main/java/org/asynchttp
>>>>>>> client/providers/grizzly/bodyhandler/PartsBodyHandler.java
>>>>>>>
>>>>>>>
>>>>>>> So, is there a way to patch AHC to stream the file so that I could
>>>>>>> eventually consume only 20mo of heap while uploading a 500mo file?
>>>>>>> Or is this simply impossible with Grizzly?
>>>>>>> I didn't notice anything related to that in the documentation.
>>>>>>> It's possible with the FeedableBodyGenerator. But if you're tied to
>>>>>>> using Multipart uploads, you'd have to convert the multipart data to
>>>>>>> Buffers manually and send using the FeedableBodyGenerator.
>>>>>>> I'll take a closer look to see if this area can be improved.
>>>>>>>
>>>>>>>
>>>>>>> Btw in my case it is a file upload. I receive a file with CXF and have
>>>>>>> to transmit it to a storage server (like S3). CXF doesn't consume memory
>>>>>>> bevause it is streaming the large fle uploads to the file system, and
>>>>>>> then provides an input stream on that file.
>>>>>>>
>>>>>>> Thanks
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>
>>>>>
>>
>