[ https://issues.apache.org/jira/browse/HTTPCORE-433?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]
Sergiu Prodan updated HTTPCORE-433: ----------------------------------- Description: This issue was observed after upgrading to httpclient v4.5.2 and httpcore v4.4.4. When configuring the PoolingHttpClientConnectionManager, I have set validateAfterInactivity to 1ms as this is the only way of maintaining the old behaviour, i.e. checking every connection if is stale before using it. I have observed a performance degradation under high load when the httpclient is shared between multiple threads. This httpclient is used only for one route. After taking a thread dump, one thing that got my attention was several threads waiting for same ReentrantLock instance while trying to AbstractConnPool#getPoolEntryBlocking/AbstractConnPool#release. The ReentrantLock instance in question was owned by another thread performing CPool#validate. It seems to me that performing this stale check inside the region protected by this lock is unnecessary and also induces a big performance hit when using the httpclient from multiple threads. I've atached a thread dump of a test application that reproduces this behaviour. ReentrantLock in question in this thread dump is 0x0000000706e8b9a8 and is owned by Thread-4. The httpclient used in this test app is used from 10 threads and has the following settings: clientConnManager.setMaxTotal(100); clientConnManager.setDefaultMaxPerRoute(100); was: This issue was observed after upgrading to httpclient v4.5.2 and httpcore v4.4.4. When configuring the PoolingHttpClientConnectionManager, I have set validateAfterInactivity to 1ms as this is the only way of maintaining the old behaviour, i.e. checking every connection if is stale before using it. I have observed a performance degradation under high load when the httpclient is shared between multiple threads. This httpclient is used only for one route. After taking a thread dump, one thing that got my attention was several threads waiting for same ReentrantLock instance while trying to AbstractConnPool#getPoolEntryBlocking/AbstractConnPool#release. The ReentrantLock instance in question was owned by another thread performing CPool#validate. It seems to me that performing this stale check inside the region protected by this lock is unnecessary and also induces a big performance hit when using the httpclient from multiple threads. I've atached a thread dump of a test application that reproduces this behaviour. ReentrantLock in question in this thread dump is 0x0000000706e8b9a8 and is owned by Thread-4 > Setting validateAfterInactivity to 1ms increases lock contention in > AbstractConnPool > ------------------------------------------------------------------------------------ > > Key: HTTPCORE-433 > URL: https://issues.apache.org/jira/browse/HTTPCORE-433 > Project: HttpComponents HttpCore > Issue Type: Bug > Components: HttpCore > Affects Versions: 4.4.4 > Reporter: Sergiu Prodan > Attachments: httpclient_thread_dump, > threads_when_inactivity_set_to_1.png, threads_when_inactivity_set_to_2000.png > > > This issue was observed after upgrading to httpclient v4.5.2 and httpcore > v4.4.4. > When configuring the PoolingHttpClientConnectionManager, I have set > validateAfterInactivity to 1ms as this is the only way of maintaining the old > behaviour, i.e. checking every connection if is stale before using it. > I have observed a performance degradation under high load when the httpclient > is shared between multiple threads. This httpclient is used only for one > route. > After taking a thread dump, one thing that got my attention was several > threads waiting for same ReentrantLock instance while trying to > AbstractConnPool#getPoolEntryBlocking/AbstractConnPool#release. The > ReentrantLock instance in question was owned by another thread performing > CPool#validate. > It seems to me that performing this stale check inside the region protected > by this lock is unnecessary and also induces a big performance hit when using > the httpclient from multiple threads. > I've atached a thread dump of a test application that reproduces this > behaviour. ReentrantLock in question in this thread dump is > 0x0000000706e8b9a8 and is owned by Thread-4. The httpclient used in this test > app is used from 10 threads and has the following settings: > clientConnManager.setMaxTotal(100); > clientConnManager.setDefaultMaxPerRoute(100); -- This message was sent by Atlassian JIRA (v6.3.4#6332) --------------------------------------------------------------------- To unsubscribe, e-mail: dev-unsubscr...@hc.apache.org For additional commands, e-mail: dev-h...@hc.apache.org