Friday, January 31, 2014

System performance utilities show wrong CPU load on Intel processosrs when HyperThreading is enables

The interesting information is published in the

 
pages 28-29:



"Compute node CPU utilization can be measured through many different tools – top, AWR, iostat, vmstat, etc. and they all give the same number, and % CPU utilization typically averaged over a set period of time. Choose whichever tool is most convenient, but allow for Intel CPU hyper-threading.
The Intel CPUs used in all Exadata models run with two threads per CPU core. This helps to boost overall performance, but the second thread is not as powerful as the first. The operating system assumes that all threads are equal thus overstates the available CPU capacity by the operating system. We need to allow for this. Here is an approximate rule of thumb that can be used to estimate actual CPU utilization, but note that this can vary with different workloads:
∙ For CPU utilization less than 50%, multiply by 1.7.
∙ For CPU utilization over 50%, assume 85% plus (util-50%)* 0.3.

Here is a table that summarizes the effect:
Measured Utilization


Actual Utilization
10%
17%
20%
34%
30%
51%
40%
68%
50%
85%
60%
88%
70%
91%
80%
94%
90%
97%
100%
100%
 "

This information is applicable to all x86 abd x86-64 servers with HyperThreading enabled.

Enabling HT causes system statistics tools - vmstat, sar - show incorrect CPU load,
and Oracle performance tools - AWR, ASH - also show and store incorrect CPU load statistics.

Be careful !



No comments:

Post a Comment

Could not locate shrept.lst make: *** [client_sharedlib] Error 1

 Installing the quarterly database patch i got unpleasant message : " Patching component oracle.sdo, 12.2.0.1.0... Make failed to ...