mersenneforum.org  

Go Back   mersenneforum.org > New To GIMPS? Start Here! > Information & Answers

Reply
 
Thread Tools
Old 2012-06-17, 15:45   #1
Jud McCranie
 
Jud McCranie's Avatar
 
Oct 2007

2×32×5 Posts
Default Is version 27.7 slow?

I'm using Prime95 on Seventeen or Bust, doing exponents in the 23,000,000 range. I had been using version 26.x on a quad-core i7, Sandy Bridge. I think it was taking under 5 days per exponent. Status was showing 4-5 days each for the expected time.

Yesterday I upgraded to version 27.7 build 2. Now it is on pace to take 6.5 to 6.9 days each. When it starts up it says it is using the all-complex AVX FFT.

Can version 27.7 be slower?
Jud McCranie is offline   Reply With Quote
Old 2012-06-17, 15:49   #2
firejuggler
 
firejuggler's Avatar
 
Apr 2010
Over the rainbow

2·5·13·19 Posts
Default

Is your CPU an AMD? I seem to recall that AVX wasn't working well in *early* 27 version.
firejuggler is offline   Reply With Quote
Old 2012-06-17, 15:51   #3
Jud McCranie
 
Jud McCranie's Avatar
 
Oct 2007

2×32×5 Posts
Default

No, it is a quad-core Intel i7, Sandy Bridge.
Jud McCranie is offline   Reply With Quote
Old 2012-06-17, 18:04   #4
Dubslow
Basketry That Evening!
 
Dubslow's Avatar
 
"Bunslow the Bold"
Jun 2011
40<A<43 -89<O<-88

160358 Posts
Default

Hmmm... (@fj: Actually, AVX has been disabled entirely for Bulldozer, because when George did get it working, it was actually slower than SSE.)

Jud, I'd say if George hasn't responded in a day or two, then try posting this in the 27.7 thread as a bug report. (Perhaps a super mod could just move these posts?)

Just to be clear, when you say SoB, you mean you're doing PRP tests, not LL tests, right?
Dubslow is offline   Reply With Quote
Old 2012-06-17, 19:26   #5
Prime95
P90 years forever!
 
Prime95's Avatar
 
Aug 2002
Yeehaw, FL

2·3·1,193 Posts
Default

27.7 with AVX support should definitely be faster than 26.x with SSE2 support.

My only guess at this point is that hyperthread detection failed. Is this a hyperthreaded CPU? Are you running with only one thread for each worker?

As a sanity check, can you also confirm via Task Manager that there is only one prime95 running and that it has 100% CPU utilization?
Prime95 is offline   Reply With Quote
Old 2012-06-17, 23:42   #6
Jud McCranie
 
Jud McCranie's Avatar
 
Oct 2007

9010 Posts
Default

I think the Seventeen or Bust test are primality tests, not PSP, but I'm not certain.

Yes, the is only one Prime95 in the task manager and it is getting 100%.

But there may be something to the hyperthreading question. This is a quad-core hyperthreaded i7, so there are eight logical CPUs. With 26.x version, the first worker ran on 1&2, the second worker on 3&4, etc. With version 27.7, if I run only one worker, it runs on only one logical CPU. If I run two workers, each one runs on one logical CPU. If I run three or four workers, each worker uses two logical CPUs.

In addition, if I am running only one worker (which is on one logical CPU), it runs at the rate of doing a test in 3.5 days, which is twice as fast as the 6.9 days per worker when I have all four workers running (on eight logical CPUs).

Number of worker windows: 4
priority: 1
all workers
most sense
smart
CPUs to use: 2

max memory, 256MB all the time

Last fiddled with by Jud McCranie on 2012-06-17 at 23:49 Reason: move
Jud McCranie is offline   Reply With Quote
Old 2012-06-18, 00:52   #7
Jud McCranie
 
Jud McCranie's Avatar
 
Oct 2007

2×32×5 Posts
Default

The Whats New for 27.3 says "However, running on 4 cores
runs into main memory bandwidth bottleneck. Sandy Bridge-E CPUs with 4 memory channels should avoid this problem."

That sounds like it might be the problem. It is a Sandy Bridge, but it is not a Sandy Bridge-E (eight cores).

Last fiddled with by Jud McCranie on 2012-06-18 at 00:54 Reason: fix
Jud McCranie is offline   Reply With Quote
Old 2012-06-18, 00:54   #8
Dubslow
Basketry That Evening!
 
Dubslow's Avatar
 
"Bunslow the Bold"
Jun 2011
40<A<43 -89<O<-88

3×2,399 Posts
Default

Windows, right? In the Test/Worker threads menu, you can select an affinity for each worker; put them on separate cores. In conjunction with that, put "AffinityScramble2=01234567" in local.txt. That should force it to mimic the v26 behavior.

Edit/PS: The memory bandwidth problem will not make it slower than v26 by any means. Instead of 25% faster it might be 15-20% faster, but still certainly not slower.

Last fiddled with by Dubslow on 2012-06-18 at 00:55
Dubslow is offline   Reply With Quote
Old 2012-06-18, 00:59   #9
Jud McCranie
 
Jud McCranie's Avatar
 
Oct 2007

1328 Posts
Default

Right, Windows 7 64-bits.

I have four workers but eight logical CPUs because of the hyperthreading. It looks like I can assign only one CPU to a worker there. Should I run eight workers, each with its own (single) CPU for best throughput?

Last fiddled with by Jud McCranie on 2012-06-18 at 01:35 Reason: logical
Jud McCranie is offline   Reply With Quote
Old 2012-06-18, 01:29   #10
Dubslow
Basketry That Evening!
 
Dubslow's Avatar
 
"Bunslow the Bold"
Jun 2011
40<A<43 -89<O<-88

160358 Posts
Default

No, disabling hyper threading altogether would be best.

You should only need to assign one CPU to a worker, and you said you already had it set at two threads per worker. Set the workers to 1, 2, 3, and 4. The local.txt line I mentioned will do the rest.
Dubslow is offline   Reply With Quote
Old 2012-06-18, 01:34   #11
Jud McCranie
 
Jud McCranie's Avatar
 
Oct 2007

2·32·5 Posts
Default

Set the four workers to CPU 1, 2, 3, 4 - or 1, 3, 5, 7? Using 1, 2, 3, 4 would use hyperthreading.

BTW, Seventeen or Bust does use P95 for PRP tests.
Jud McCranie is offline   Reply With Quote
Reply

Thread Tools


Similar Threads
Thread Thread Starter Forum Replies Last Post
Prime95 version 26.5 built 5 slow! Unregistered Information & Answers 7 2012-06-17 05:22
Slow Down? R.D. Silverman GMP-ECM 55 2011-10-16 17:28
GMP ECM slow on my laptop? R.D. Silverman GMP-ECM 9 2011-02-17 14:57
How hot is too hot? Slow is too slow? petrw1 Hardware 13 2008-11-10 23:25
Really slow machines? Doorbasher Hardware 5 2004-08-23 22:18

All times are UTC. The time now is 14:54.

Wed Nov 25 14:54:37 UTC 2020 up 76 days, 12:05, 3 users, load averages: 1.90, 1.65, 1.51

Powered by vBulletin® Version 3.8.11
Copyright ©2000 - 2020, Jelsoft Enterprises Ltd.

This forum has received and complied with 0 (zero) government requests for information.

Permission is granted to copy, distribute and/or modify this document under the terms of the GNU Free Documentation License, Version 1.2 or any later version published by the Free Software Foundation.
A copy of the license is included in the FAQ.