topic rdtsc anomaly in Intel® Moderncode for Parallel Architectures

rdtsc anomaly

slackmack — Tue, 11 Aug 2009 04:10:39 GMT

I wrote a test program to compare the results from rdtsc and gettimeofday. I ran the test of Nehalem X5570. I understand that this platform has Invariant TSC.

The test program is as follows:
1. start = x86_get_timer();
2. while loop downcounting register value from 2.93e+09 till 0
3. end = x86_get_timer();
4. Display the result.

With just the above steps, the result printed varying drastically. I can't explain why it is so, even when I do taskset to run the program.

I added gettimeofday calls before and after to compare the results. The results of both rdtsc and gettimeofday were consistent, but the result of gettimeofday is almost half of rdtsc value.

I've attached the program for readers to review and comment.

Thanks,
SlackMack

Re: rdtsc anomaly

jimdempseyatthecove — Tue, 11 Aug 2009 12:28:34 GMT

Try

[cpp]    return ((((unsigned long long)hi)<<32) | (lo & 0xffffffff));[/cpp]

Jim Dempsey

Re: rdtsc anomaly

TimP — Wed, 12 Aug 2009 01:01:57 GMT

Quoting - jimdempseyatthecove

Try

[cpp]    return ((((unsigned long long)hi)<<32) | (lo & 0xffffffff));[/cpp]

Jim Dempsey

With Intel or Microsoft compilers, the _rdtsc() macro takes care of this stuff, and the return value may be treated as unsigned long long int.

rdtsc anomaly

zhen436 — Mon, 08 Mar 2010 17:23:58 GMT

Does the suggested solution work?

rdtsc anomaly

TimP — Mon, 08 Mar 2010 18:37:08 GMT

Neither you nor the original poster specified which compiler you intend to use, or which defines were set, so the reasonable assumption was made that correcting the obvious error in combining low and high order parts would help.