Weblogic.socket.PosixSocketMuxer.poll() take 25% of CPU into jvisualvm

Question

Hello,
I'm running a bench of my software and I reach the CPU limit of my 2 application servers. When I'm looking for what method is the most consuming using jvisualvm sampler plugin, I can see the
weblogic.socket.PosixSocketMuxer.poll() method take 25% of CPU.
I used to run some benches under Solaris, and I didn't see the DevPollSocketMuxer consuming like that.
And obvisouly, my servers are strong : 2 x3650M2 with 2xX5560 quad cores 2 threads (16 virtual CPU per servers)
I'm using weblogic 10.3.4.
Could you please help me to give me some investigation tracks ?
Best Regards,

Answer 1

Something spring to my mind when reading your initial post.
"And obvisouly, my servers are strong : 2 x3650M2 with 2xX5560 quad cores 2 threads (16 virtual CPU per servers)"
How many threads of the weblogic.socket.Muxer (weblogic.socket.PosixSocketMuxer) are there?
The native I/O muxer uses its own thread pool, i.e., it uses N + 1 threads by default, where N is the number of CPUs.
When running on multi-core CPUs that use hyper threading, it has an impact on the CPU count. In most cases, it makes
sense to reduce the number of muxer threads to eliminate the double counting of CPUs caused by hyper-threading.
In your case (2xX5560 quad cores 2 threads) the number of muxer threads would be 2 * 4 * 2 + 1 = 17 threads
Too many native muxer threads can have an impact on performance and it very well could be that this is what you
are observing "and I reach the CPU limit of my 2 application servers". You can change the number of native CPU
threads by using the weblogic.SocketReaders system property and set it to 2 + 1 = 3 threads.