Detailed client software thread... - Page 3

adpowers · 2003-02-11, 08:08

Thank you. I finally got around to installing it and it gave me a .001 iteration improvement on a normal LL test on my P4 :).

Jwb52z · 2003-02-11, 19:13

On this board, for me, the ftp links don't work. On my last computer they did. All they do now is just open a new white blank window where the XP flag just waves up in the corner of the browser. I have no idea what is wrong.

cheesehead · 2003-02-11, 19:23

Quote:

Originally Posted by Jwb52z

On this board, for me, the ftp links don't work. On my last computer they did. All they do now is just open a new white blank window where the XP flag just waves up in the corner of the browser. I have no idea what is wrong.

It's just that the mersenne.org server is having trouble right now. For over an hour I haven't been able to get to any mersenne.org page.

Try again later.

Prime95 · 2003-02-20, 18:19

The wait is over.... P4 owners can expect a 4.5% to 6% performance increase over version 23.1 in LL testing on all FFT sizes!

Windows: ftp://mersenne.org/gimps/p95v232.zip
Linux: ftp://mersenne.org/gimps/mprime232.tar.gz
ftp://mersenne.org/gimps/sprime232.tar.gz
NT service: ftp://mersenne.org/gimps/winnt232.zip

Let me know if you find any problems. Overclockers should keep an eye out for hardware errors. I haven't run into any troubles on my own overclocked machine, but the increased performance *might* generate more heat or require more voltage or require a 1MHz reduction in clock speed. Happy crunching!

ewmayer · 2003-02-20, 22:33

Hey Gorge

Can you give us a clue as to what you did to get the performance boost?
This isn't the better-use-of-L2-cache you mentioned recently, is it?

Ian_H · 2003-02-20, 23:06

Thanks George!

Looks like around 4.5% faster on a P4 1.7 here at home. Will install tomorrow on 1.7 and 1.8 P4s at work. P4 2.0 coming next week...

Prime95 · 2003-02-21, 01:31

Quote:

Originally Posted by ewmayer

Can you give us a clue as to what you did to get the performance boost?

I optimized each building block macro. A building block is something that takes data from 2 or 4 cache lines and does something - like take four complex numbers and do two levels of the FFT. This basically entails trying a variety of different implementations, moving independent sections around, etc. looking for the implementation that is the quickest. These small macros generally take 80 to 200 clocks and in most cases I was able to save anywhere between zero and 15 clocks. I think some could be improved further, but I've not been able to hit on the magic sequence of instructions. While I have no hard and fast rules to pass on it seems that increasing independence is good (e.g. (A + B) + (C + D) is better than ((A+ B) + C) + D) as long as it does not increase memory accesses above a certain threshold.

Next up, reducing TLB thrashing by copying FFT data in pass 1 to a contiguous memory block.

P.S. The beauty of SSE2 coding as opposed to old x86 coding is that directly addressable registers let you can move an instruction up or down compile and test. In the stack based x86 world moving an instruction requires rewriting the rest of the macro because you've changed the order of the data item on the FP stack. In practice this means you can try something in two minutes that would require an hour to try on the x86.

Ian_H · 2003-02-21, 17:14

Follow up after 18 hours on the P4 1.7: steady and consistent speed increase of over 13 per cent! Iterations previously 0.77 (or 0.76 very occasionally), now constant 0.68. (Exponent 15710239)

How come this is so much more of a speed increase than you predicted?

QuintLeo · 2003-02-22, 01:18

P4 2.0a oc 119/478 FSB, mem 4/3 of FSB to 158ish.
Slackware 8.1 / 2.4.18 custom kernel, SS51G box.

Went from .98 (22.9) to .93 (23.1) to .88 (23.2 benchmarks, haven't been running it long enough to see actual timings in practice yet).
No other changes made except replacing mprime with the newer version.

Nice little improvement, makes me wish I had more SSE2 boxes to work with. Come Clawhammer time....

9-)

rectifire · 2003-02-22, 02:35

I have just tried out the new prime95 client. I've calculated that it brought me an almost 14% speed increase over version 22.12 Very nice. Thanks a lot!

I'm using a p4 2.4Ghz with DDR333 RAM

QuintLeo · 2003-02-23, 01:53

Update - actual run is mostly .089, with an occasional .088

Still a nice 10% or a bit more improvement over 22.9

2003-02-20, 18:19	#26
Prime95 P90 years forever! Aug 2002 Yeehaw, FL 3⁵×31 Posts	Version 23.2 The wait is over.... P4 owners can expect a 4.5% to 6% performance increase over version 23.1 in LL testing on all FFT sizes! Windows: ftp://mersenne.org/gimps/p95v232.zip Linux: ftp://mersenne.org/gimps/mprime232.tar.gz ftp://mersenne.org/gimps/sprime232.tar.gz NT service: ftp://mersenne.org/gimps/winnt232.zip Let me know if you find any problems. Overclockers should keep an eye out for hardware errors. I haven't run into any troubles on my own overclocked machine, but the increased performance might generate more heat or require more voltage or require a 1MHz reduction in clock speed. Happy crunching!

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
Program for detailed analysis of ECM-NFS crossovers	Dubslow	Factoring	14	2016-03-18 06:01
Detailed Stats 0 to 100 Gauges	tapion64	PrimeNet	2	2014-04-12 08:26
More Detailed Exponent/Assignment Status	KingKurly	PrimeNet	8	2011-01-06 20:51
linux 25.9 client, unsafe thread output?	xorbe	Software	0	2009-04-03 04:21
"Latest client software" missing v25.6	OneOfMany	Software	1	2008-09-18 16:48

2003-02-11, 08:08	#23
adpowers Sep 2002 2⁴×5 Posts	Thank you. I finally got around to installing it and it gave me a .001 iteration improvement on a normal LL test on my P4 :).

2003-02-11, 19:13	#24
Jwb52z Sep 2002 17·47 Posts	On this board, for me, the ftp links don't work. On my last computer they did. All they do now is just open a new white blank window where the XP flag just waves up in the corner of the browser. I have no idea what is wrong.

2003-02-20, 22:33	#27
ewmayer ∂²ω=0 Sep 2002 República de California 2×3²×647 Posts	Hey Gorge Can you give us a clue as to what you did to get the performance boost? This isn't the better-use-of-L2-cache you mentioned recently, is it?

2003-02-20, 23:06	#28
Ian_H Aug 2002 Europe 2·3·5 Posts	Thanks George! Looks like around 4.5% faster on a P4 1.7 here at home. Will install tomorrow on 1.7 and 1.8 P4s at work. P4 2.0 coming next week...

2003-02-21, 17:14	#30
Ian_H Aug 2002 Europe 2·3·5 Posts	Follow up after 18 hours on the P4 1.7: steady and consistent speed increase of over 13 per cent! Iterations previously 0.77 (or 0.76 very occasionally), now constant 0.68. (Exponent 15710239) How come this is so much more of a speed increase than you predicted?

2003-02-22, 01:18	#31
QuintLeo Oct 2002 Lost in the hills of Iowa 2⁶×7 Posts	P4 2.0a oc 119/478 FSB, mem 4/3 of FSB to 158ish. Slackware 8.1 / 2.4.18 custom kernel, SS51G box. Went from .98 (22.9) to .93 (23.1) to .88 (23.2 benchmarks, haven't been running it long enough to see actual timings in practice yet). No other changes made except replacing mprime with the newer version. Nice little improvement, makes me wish I had more SSE2 boxes to work with. Come Clawhammer time.... 9-)

2003-02-22, 02:35	#32
rectifire Feb 2003 2₈ Posts	I have just tried out the new prime95 client. I've calculated that it brought me an almost 14% speed increase over version 22.12 Very nice. Thanks a lot! I'm using a p4 2.4Ghz with DDR333 RAM

2003-02-23, 01:53	#33
QuintLeo Oct 2002 Lost in the hills of Iowa 2⁶×7 Posts	Update - actual run is mostly .089, with an occasional .088 Still a nice 10% or a bit more improvement over 22.9