topic Constant execution time on P4? in Intel® Integrated Performance Primitives

Constant execution time on P4?

arintel1 — Mon, 07 Mar 2005 15:27:12 GMT

Dear friends,

In theory, can we achieve constant execution time on Pentium 4 (with and without Hyper-Threading Technology) ? How?

Sincerely,

ARINTEL

Re: Constant execution time on P4?

Vladimir_Dudnik — Thu, 10 Mar 2005 22:27:46 GMT

Hi,

what do you mean? Constant execution time for some program, which will be the same each time it runned? With clock-to-clock accuracy?

Regards,

Vladimir

Re: Constant execution time on P4?

arintel1 — Fri, 11 Mar 2005 00:03:24 GMT

Hi,

That's exactly what I meant.

I tried the following things with the routine that I wanted to have constant execution time:
1. Put it in kernel mode of Windows XP running at the highest IRQL (31) -> only be interrupted by NMI
2. The routine contains a loop which repeats 10 times the same task: processing data and measuring CPU clock cycle. I take the last measurement as the final result.
3. On P4 with Hyperthreading, I made the second logical processor get stuck in a spin loop while the other processor was processing the routine.

With those things, I could achieve 99.99% constant execution time for the routine on all Celerons and Pentiums except Pentium 4 (with or without HyperThreading).

On P4,the measurement result constantly fluctuates in some range, though small.

Thanks!

Re: Constant execution time on P4?

Vladimir_Dudnik — Sat, 12 Mar 2005 03:09:05 GMT

Hi,

there are might be many reasons for that. For example hardware prefetcher on Pentium 4 can optimize memory accesses after it accumulate enough history.

Regards,

Vladimir