The "Big Iron Test"

Posted by Benny Tritsch on November 25, 2004 – updated on January 31, 2005

[Introduction] [Test Setup] [Testing Methodology] [Analysis 1] [Analysis 2] [Analysis 3]

Analysis of Test 2: Windows Server 2003 with HyperThreading turned off

The second test was done on the same hardware platform with eight physical CPUs, but with HyperThreading turned off. This time, it only took just over one hour until 100 users were logged in. Again the Terminal Server system showed increasing symptoms of heavy load, but not as significant as during the first test. It took about 40 more minutes until an additional 30 users were fully logged in.

 

 

However, the CPU load still showed a permanent load of more than 20% for all CPUs at the end of the test. Memory consumption was at about the same level as during the previous test and the Processor Queue Length was even higher; it came to an average of 45 by the end of this test.

 

 

With 130 users logged in, this high-end machine was getting very slow. The test users were not interacting with the applications at any time during the test; the applications were simply occupying memory. System-internal administrative tasks thus accounted for the full processor load. System reaction time to user actions dropped noticeably from the time previously measured. Simple dialog and window changes took more than 10 seconds.

At this point we decided to take a closer look at the monitored values of the Context Switches per second. This performance counter measures the combined rate at which all processors on the computer are switched from one thread to another. Context Switches occur when a running thread voluntarily relinquishes the processor, is preempted by a higher priority ready thread, or switches between user-mode and privileged (kernel) mode to use an Executive or subsystem service. It is the sum of Thread\Context Switches per second for all threads running on all processors in the computer and is measured in numbers of switches. There are Context Switch counters on the system and thread objects. This counter displays the difference between the values observed in the last two samples, divided by the duration of the sample interval.

 

 

The result was a major surprise: more than 40,000 Context Switches per second in average by the end of the test. The CPUs were switching from one thread to another and synchronizing their activities.

Immediately, prior to the end of the test, the Windows Task Manager showed that the load was not balanced equally between the eight CPUs. It was obvious that adding more CPUs could make the problem even worse.

 

 

Next