Detailed client software thread...

[db597]

Cheers for that. Now running on 23.4. You've done a great job! :)

How much more juice do you think you can squeeze out of SSE2? We're a full 20% faster than the 22.x clients... how much more is possible?

[Prime95]

Originally Posted by db597

How much more juice do you think you can squeeze out of SSE2? We're a full 20% faster than the 22.x clients... how much more is possible?

I'm running out of ideas and I've implemented the easy ones. There are still a few possibilities...

Last night I discovered a way to speed up the normalization code (2 fewer multiplies), but it will be hard to implement. This speeds up timings by 1-2%.

More FFT sizes. Also not easy. Saves roughly 5-10% for about half of the exponents.

More use of data while in the L1 cache. This is what I did in 23.4. Unfortunately, doing more of this will be a bit of work and may or may not speed up prime95.

[Daffy]

Originally Posted by Prime95

More improvements for P4 owners! You can expect up to 9% faster iteration times compared to version 23.3! All users doing ECM should also upgrade to 23.4.

Just for curiosity, do you keep track of the versions being used by participants ? And do you have statistics ?

Since Prime95 has evolved quite a bit, I was wondering how everyone is keeping up to date.

[db597]

More FFT sizes. Also not easy. Saves roughly 5-10% for about half of the exponents.

There is some computing time wasted for exponents near the filter cut off - since results are not the same as expected, some parts are rerun to see if the error is reproduceable. Anyway to reduce this "retesting"?

Or would having more FFT sizes mean being near the filter cut off more often? Hence, having more "retesting"?

[db597]

Just for curiosity, do you keep track of the versions being used by participants ? And do you have statistics ?

Since Prime95 has evolved quite a bit, I was wondering how everyone is keeping up to date.

The mersenne.org front page and download page are still advertising 22.12 clients. :( It's likely that many people don't even know about the latest code.

[cheesehead]

Originally Posted by db597

There is some computing time wasted for exponents near the filter cut off - since results are not the same as expected, some parts are rerun to see if the error is reproduceable. Anyway to reduce this "retesting"?

I don't know how many iterations are run while testing for good results near FFT changeover points, but I'll bet it's not much more than 10,000. And they needn't be re-run unless there is a switch to the larger FFT size.

If you're L-Ling an exponent near 19,999,000, for instance, the total number of iterations for the L-L will be about 19,999,000. 10,000 is about 0.05% of 19,999,000, not a very noticeable fraction compared to a 5% savings from changing something else.

You could take the square root of the total number of iterations to find number of iterations you should test. For 19999000 iterations that would be 4472.

This technique is used for gallups too.

[Prime95]

Originally Posted by db597

There is some computing time wasted for exponents near the filter cut off - since results are not the same as expected, some parts are rerun to see if the error is reproduceable. Anyway to reduce this "retesting"?

I think you are talking about the retesting when a roundoff > 0.40 is found. If your save files are written every half hour (the default) the average wasted time is 15 minutes per error. If you get 10 errors during the LL test, thats 150 minutes wasted. That should be less than the time spent using a larger FFT size.

You can reduce your time between save file writes. This would reduce wasted time retesting, but would waste extra time writing more save files.

[nomadicus]

Originally Posted by Daffy

Since Prime95 has evolved quite a bit, I was wondering how everyone is keeping up to date.

Seems like we need to get the mersenne.org home page updated with V23.4 instead of V22. Peronally I wouldn't know how to get V23.4 if it wasn't for this forum.

[S80780]

Just take a look at ftp://mersenne.org/gimps/. There, you'll find several versions of Prime95/mprime/prp, sourcecodes, statistics ...

I thing the main reason for the homepage still offering v22 is that this version has a better performance on non-P4 Systems.

Benjamin

[db597]

I think you are talking about the retesting when a roundoff > 0.40 is found. If your save files are written every half hour (the default) the average wasted time is 15 minutes per error. If you get 10 errors during the LL test, thats 150 minutes wasted. That should be less than the time spent using a larger FFT size.

Yep! That's the one. Is there any way to reduce that 150 wasted minutes?