I run a test in my server with 4 socket server,there are so much remote memory access , so I want to figure out whether the QPI traffic be the bottleneck. I run pcm to monitor QPI outgoing traffic and find out that sometime QPI port outgoing traffic can be 22G. as I know ,the QPI frequency is 9.6GT/s , so the QPI speed is 19.2GB/s, why the traffic can be 22G per second larger than 19.2G ?
This is strange. What QPI speed is PCM detecting? This is reported somewhere in the header of the output. Is PCM reporting that it can't read teh QPI_STATUS_REGISTER? If so, the QPI speed is measured emprically, which is subject to errors of measurements.
Finally, what does PCM report as QPI link utilization? The QPI bandwidth is computed based on the QPI utilization.
thank you for your reply.
the QPI speed is 19.2GBytes/second (9.6GT/second). there are not reporting that it can't read teh QPI_STATUS_REGISTER.
The QPI Link utilization is about 75% while the outgoing traffic from this qpi port is 22G. And sometime it can be 77% while traffic is 20G.
it seem the 'max bandwidth' changes all the time ...