P-1 factoring anyone?

[Kevin]

Just noticed this thread now, so I'm going to put a few cores on P-1. Is there a nice/preferred way to split up work to prevent resource bottlenecks? I've got a Q9550 w/64 bit Linux and a Q6600w/ 32 bit Vista, both with 4GB 1066mhz memory. I don't particularly mind lower system performance if it's for a good cause .

[Prime95]

Originally Posted by Kevin

Just noticed this thread now, so I'm going to put a few cores on P-1. Is there a nice/preferred way to split up work to prevent resource bottlenecks? I've got a Q9550 w/64 bit Linux and a Q6600w/ 32 bit Vista, both with 4GB 1066mhz memory. I don't particularly mind lower system performance if it's for a good cause .

Giving prime95 1 or 1.5 GB should be enough for 2 P-1 workers.

[Andi47]

On one of the computers which I have access to, I have installed P95, version 25.8 build 5, and set it to trial factoring. A few days later I changed the setting for preferred work type to p-1 factoring, and changed the day/nighttime memory to 128 / 512 MB. When I checked back now, I am surprised that it got a LL double-check assigned. Why this? Is 512 MB nighttime memory not enough to do stage 2? (increased it to 768 MB now.)

[James Heinrich]

Originally Posted by Kevin

Is there a nice/preferred way to split up work to prevent resource bottlenecks?

For a dedicated P-1 machine I'd suggest two 2-threaded P-1 workers, with MaxHighMemWorkers=1 set in local.txt, and ~2000MB allocated (more, if you're feeling generous, but 2G is plenty for any current P-1 job). This way you perpetually have 1 worker using all the memory for stage2, and the other worker working on stage1 -- you'll never have any "wasted" time where there is not a lot of RAM being used. There is a chance, however, that this setup would result in stage1 workers running out of things to do if stage2 takes a bit longer to run than stage1.

Running 4 P-1 workers (allowing 2 of them to do stage2 at any given time) works, but may choke your system more than 2 workers (but avoids the stage1-starvation problem).

You could have two P-1 workers, single-threaded, with the other 2 threads working on other less-memory-hungry work like TF.

[Mr. P-1]

Originally Posted by James Heinrich

For a dedicated P-1 machine I'd suggest two 2-threaded P-1 workers, with MaxHighMemWorkers=1 set in local.txt, and ~2000MB allocated (more, if you're feeling generous, but 2G is plenty for any current P-1 job). This way you perpetually have 1 worker using all the memory for stage2, and the other worker working on stage1 -- you'll never have any "wasted" time where there is not a lot of RAM being used. There is a chance, however, that this setup would result in stage1 workers running out of things to do if stage2 takes a bit longer to run than stage1.

In my experience, stage2 takes a LOT longer to run than stage1, so that a backlog of uncompleted stage2 work would be rapidly built up using this strategy. That's not a problem for me, because I'm willing to set MaxHighMemWorkers to 2 periodically in order to clear it.

An alternative for a 4 core machine would be to have three cores doing P-1 with MaxHighMemWorkers=2

Or you could have 2 cores doing P-1, and allow both to use high memory at night, but only one during the day. See undoc.txt for how to do this.

[Mr. P-1]

Originally Posted by Andi47

On one of the computers which I have access to, I have installed P95, version 25.8 build 5, and set it to trial factoring. A few days later I changed the setting for preferred work type to p-1 factoring, and changed the day/nighttime memory to 128 / 512 MB. When I checked back now, I am surprised that it got a LL double-check assigned. Why this? Is 512 MB nighttime memory not enough to do stage 2? (increased it to 768 MB now.)

512MB is ample for one P-1 thread, and sufficient, albeit a bit minimal, for two. You probably need to have nighttime memory available for about 2/3 of the total time your computer is on. Otherwise you may need to allow stage2 to run during the day.

I don't know why you got the LL assignment.

[Mr. P-1]

Originally Posted by Prime95

Giving prime95 1 or 1.5 GB should be enough for 2 P-1 workers.

I find it a bit unintelligent in how it allocates memory when there are two threads competing for it. Specifically if one thread only has a small number of relative primes to finish, the other thread will grab all the memory, leaving the first with a small amount of memory to start the next stage2.

This contrasts with the situation where the second thread finishes stage 1 and wants to continue with stage2. In this case, the first thread is restarted, with all the overhead of reinitialising the temporary variables that this entails.

Better, perhaps, in both of these situations, would be to leave the second thread doing low memory work until the first finishes its current pass. Then both could start with half the available memory.

[petrw1]

Originally Posted by James Heinrich

For a dedicated P-1 machine I'd suggest two 2-threaded P-1 workers, with MaxHighMemWorkers=1 set in local.txt, and ~2000MB allocated (more, if you're feeling generous, but 2G is plenty for any current P-1 job). This way you perpetually have 1 worker using all the memory for stage2, and the other worker working on stage1 -- you'll never have any "wasted" time where there is not a lot of RAM being used. There is a chance, however, that this setup would result in stage1 workers running out of things to do if stage2 takes a bit longer to run than stage1.

Running 4 P-1 workers (allowing 2 of them to do stage2 at any given time) works, but may choke your system more than 2 workers (but avoids the stage1-starvation problem).

You could have two P-1 workers, single-threaded, with the other 2 threads working on other less-memory-hungry work like TF.

On my Q9550 I have 3 workers on LL and 1 on P-1 with 1200Mb.
I find Stage 1 takes 21 hours and Stage 2 takes 37 hours.

However, the BAD news is that when I compare P-1 alone against P-1 with 3 LL workers the P-1 stage 2 runs about 60% slower with the 3 LL's running. I didn't test Phase 1 yet.

[Kevin]

Originally Posted by petrw1

On my Q9550 I have 3 workers on LL and 1 on P-1 with 1200Mb.
I find Stage 1 takes 21 hours and Stage 2 takes 37 hours.

However, the BAD news is that when I compare P-1 alone against P-1 with 3 LL workers the P-1 stage 2 runs about 60% slower with the 3 LL's running. I didn't test Phase 1 yet.

Have you tried running P-1 and trial factoring on one pair of cores, and two LL tests on the other pair? I've got a fairly sizable cache of sub-33M first time tests I'm working through, so I'd like to keep my quads at least 50% on LL testing. If there's no "good" way to run P-1 and LL simultaneously, then I'll probably keep the quads 100% on LL testing and switch them over to P-1 once I cut down the cache considerably.

[petrw1]

Originally Posted by Kevin

Have you tried running P-1 and trial factoring on one pair of cores, and two LL tests on the other pair? I've got a fairly sizable cache of sub-33M first time tests I'm working through, so I'd like to keep my quads at least 50% on LL testing. If there's no "good" way to run P-1 and LL simultaneously, then I'll probably keep the quads 100% on LL testing and switch them over to P-1 once I cut down the cache considerably.

I also had a suggestion to try DC instead of LL, implying a smaller FFT might help. I might try both.

The good news is that PM-1 does NOT impact LL much and like you my personal focus is to clean up my own 32 sub-33M LL tests and get my LL stats up.

See end of this post: http://www.mersenneforum.org/showpos...4&postcount=54

[Andi47]

On a PC which has got 1 GB Ram installed, I had a memory setting of 256 MB daytime (7:30 to 19:30) and 786 MB at night. When it requested an assignment during daytime (preferred type is set to p-1) it said "not enough memory for p-1 factoring assignments".

Does P95 check for day/nighttime memory settings when trying to get p-1 or ecm assignments?

I have now changed my memory settings (added a line in prime.txt), with this settings it should get max. memory in 68% of the time in a week:

Memory=900 during 1-5/18:00-24:00,1-5/0:00-7:15,6-7/0:00-24:00 else 256

would this be enough to get p-1 type assignments?

BTW: Is it possible to specify certain days of the year where Prime95 is allowd to take max. memory? (for example: nobody is in office and using the PC on holidays like christmas.)