Re: QueryPerformanceCounter question

Tech-Archive recommends: Repair Windows Errors & Optimize Windows Performance

Thanks Daniel,

If you're keen to make a framework, you'll probably get the most accurate
and stable results by using QPC with thread priority boosted all the way
up and thread affinity set to a single CPU. Of course, this is only
practial in a test environment where you're assessing the performance of
individual functions or other compute-bound workloads one against another.
You'd never want to attempt to measure the performance of a real-world
application using such techniques, and forget about measuring the
performance of multiple threads simultaneously using this technique (well,
maybe 1 thread per CPU, but even that's probably pushing it).

You can use GetThreadTimes, but the resolution is very low. You're likely
to see a lot of aliasing in your results unless your threads remain
compute bound for long periods of time, or you gather data over a long
period of time (several minutes at least - maybe more).

Thanks pretty much the conclusion I came to last night night.

"Carl Daniel [VC++ MVP]" <cpdaniel_remove_this_and_nospam@xxxxxxxxxxxxxxx>
wrote in message news:OtsJwQ5rGHA.4856@xxxxxxxxxxxxxxxxxxxxxxx
DC wrote:
Ah. I still haven't had a chance to look at the code.

In response to the issues you raised, I guess I can use a thread
affinity mask to force the scheduler into keeping me on 1 cpu?

The thing is, I still don't have much more to go on as to what the
best approach to profiling is. Bear in mind, I wanted to be as
accurate as possible and if a thread context switch occurs between
the 2 QPC calls then my result will be affected. Worse, when I run
the same (of different) test again a context switch may not happen.
This is why I was looking at GetThreadTimes.

What's the opinions on the this?

The best way to do profiling is to use a profiler.

The fact is, that running under a preemptive OS like windows, there's
simply no way to measure performance independently from other activity on
the system. The closest that you can practially get for most real-world
apps is to use a sampling profiler to gather statistical profile data for
your application while running a test workload over a relatively long
period of time. The sampling profiler removes a great deal of the
influence of the rest of the system since only samples where the CPU was
actually executing your process/threads are counted.

If you're keen to make a framework, you'll probably get the most accurate
and stable results by using QPC with thread priority boosted all the way
up and thread affinity set to a single CPU. Of course, this is only
practial in a test environment where you're assessing the performance of
individual functions or other compute-bound workloads one against another.
You'd never want to attempt to measure the performance of a real-world
application using such techniques, and forget about measuring the
performance of multiple threads simultaneously using this technique (well,
maybe 1 thread per CPU, but even that's probably pushing it).

You can use GetThreadTimes, but the resolution is very low. You're likely
to see a lot of aliasing in your results unless your threads remain
compute bound for long periods of time, or you gather data over a long
period of time (several minutes at least - maybe more).

-cd




.

Relevant Pages

  • Re: Per databse CPU usage
    ... You might want to use a server-side trace instead of Profiler if you're ... SQL Server MVP ... > We are experiencing CPU performance issues, some requests just> take all the CPU for 10 to 20 seconds. ...
    (microsoft.public.sqlserver.server)
  • Re: Per databse CPU usage
    ... This article is a pretty> good intro, and the techniques also apply to extracting data from trace> tables saved via Profiler: ... > SQL Server MVP ... >> I'm currently trying to track down these heavy requests to determine ... >> which databses are using CPU the most, ...
    (microsoft.public.sqlserver.server)
  • Re: Enforce (MTS) & Performance
    ... I'm sorry Andrew, I got too busy with the problem..... ... I have the sample Profiler output. ... 1953 in the CPU column ?? ... >> Stored Proc with a cursor that executes another Stored ...
    (microsoft.public.sqlserver.server)
  • Re: QueryPerformanceCounter question
    ... affinity mask to force the scheduler into keeping me on 1 cpu? ... The best way to do profiling is to use a profiler. ... application while running a test workload over a relatively long period of ... rest of the system since only samples where the CPU was actually executing ...
    (microsoft.public.win32.programmer.kernel)
  • High CPU Utilization on Periodic basis
    ... I used sql profiler over a period of time to see which sps ... were cpu hogs during the normal periods and the high cpu ... >type, since the graphs show a sudden spike, not a ramp up. ...
    (microsoft.public.sqlserver.server)