Version 28.7 slower?
 

Hi,

I just upgraded to version 28.7 of the software. I apologize for not knowing what I had previously. In any case, the LL test I am running (> 100,000,000 digit) is running slower than it did on the older version. I was getting about 48 ms/iteration on my Core i7-3740QM CPU, and now I get about 52 ms/iteration, running 4 worker threads on the one test.

Any idea why this is? Should I install an older version? I saw that a previous version had a bug in it...I'm pretty sure what I was using previously was prior to that. Anyway, what's the last release that doesn't have a bug?

Thanks,
Drew

Compared to which version? You might can figure this out from looking at the file dates at
[url]ftp://mersenne.org/gimps[/url]

You can plow through the whatsnew.txt and see if any of the comments might apply to your situation. But, in general, I cannot remember any changes that should have affected you to any significant degree.

Also, you might try 28.9 when it comes out as there are more subtle changes.

There shouldn't be any problem with re-installing an older version should you decide to go that route.

[QUOTE=drew;430227]Hi,

I just upgraded to version 28.7 of the software. I apologize for not knowing what I had previously. In any case, the LL test I am running (> 100,000,000 digit) is running slower than it did on the older version. I was getting about 48 ms/iteration on my Core i7-3740QM CPU, and now I get about 52 ms/iteration, running 4 worker threads on the one test.

Any idea why this is? Should I install an older version? I saw that a previous version had a bug in it...I'm pretty sure what I was using previously was prior to that. Anyway, what's the last release that doesn't have a bug?
[/QUOTE]

Is it possible (George may have more insight) that the new version is setting affinity differently than the older one? With 4 threads on the single worker, if it's doing something like accidentally using hyperthreads instead of actual cores, that could cause some issues. Although probably more than an 8.3% slowdown.

I'm assuming it's running the test at the same FFT size as before, but could be worth checking.