Separate names with a comma.
Discussion in 'Distributed Computing' started by Orange Kid, Mar 2, 2017.
Well done, TeAm mates!
Congrats to you guys, wish I would have set my stuff up for it before I worked the weekend. Looks like I am going to pull out some of the systems I just retired in March for these CPU only tasks. Its not much, some Core2Duo's and a quad AMD A8, but I'll have them ready to roll. Just have to find some space, garage is too hot now and my daughters room, where I have the other systems, has to get reorganized before she comes home for summer break from school in 3 weeks.
"Never mind that whirring sound coming from under the bed, honey!"
the "bad" ??, you mean the bed ?
Finally! The medals engraver was slow this time around, but here they are:
Now you have me hooked on CPU races.... When is the next ?
Still can't believe we beat OCN, thanks to you, @Markfw !
We don't know what the next sprint will consist of, it will be announced 4/26 at 2:00PM Pacific. I don't think I will be able to do much, since my vacation is from 4/27 to 5/3, but I will try to control things from my laptop...
There's also a PrimeGrid CPU race going. Best run on Intel CPUs with AVX.
Is primegrid part of BOINC ? Doesn't Ryzen have AVX ? I have 112 thread on 4 E5-2683 Xeons, and those would work, and 2 Ryzen that should work, plus a few other misc Intel.
And how long is that race ? WCG is my primary CPU DC app. That and F@H are the best cancer research I could find, and thats my goal.
the Ryzen 8 cores are beating the E5's in some apps. Even with only 8 cores vs the E5 at 14. Amd yes Ryzen supports AVX
Per this page: https://www.primegrid.com/forum_thread.php?id=7323
The Ryzen beats Haswel, Broadwell, and barely, Skylake on Primegrid.
The PrimeGrid race ends in about 4.5 days. Llr, which is being used for this PrimeGrid race, is heavily optimized for Intel FPUs. Ryzen might get one or two WUs done in time if you enable parallel computation with app_config.xml. See the PrimeGrid races thread.
I just added Primegrid to one box, My Ryzen 1800x just to benchmark it. Added at 9 pm PST to Anandtech team. If someone can get me a link to track its progress, we can see how it does compared to other CPU's
OK, I looked, but its a little confusing. To speed up this effort, can you give me a real example for a 8 core 16 thread app_config.xml ?
I see this:
<app_config> <app> <name>llrSOB</name> <fraction_done_exact/> </app> <app_version> <app_name>llrSOB</app_name> <cmdline>-t 4</cmdline> <avg_ncpus>4</avg_ncpus> <max_ncpus>4</max_ncpus> </app_version> </app_config>
So would I change all the 4's to 8 ? or 16 ?
And where is that file ? I can't find it
OK, I found your computer. http://www.primegrid.com/results.php?hostid=870074
It's doing the wrong tasks for this race. PrimeGrid has a wall'o'options. You need to set only Prime Sierpinski Problem LLR for the CPU, and disable both GPU checkboxes at the top. http://www.primegrid.com/prefs.php?subset=project
How do I set those tasks ? I have never done that before.
Edit, and GPU is disabled under activity.
Edit: I may have found the "Prime Sierpinski Problem" choice, and I aborted the other tasks, but now I see what you mean on using more CPU's on one task, but I need the location and content of that config file.
Click the "Edit PrimeGrid preferences" link. Then uncheck the second and third checkboxes, uncheck your sieve checkbox in the CPU column, and check the check box for Prime Sierpinski Problem LLR (PSP). Be sure to save the changes.
Create a file called app_config.xml in c:\application data\BOINC\projects\www.PrimeGrid.com . (I think that's right.) Save this to the file:
<app_config> <app> <name>llrPSP</name> <fraction_done_exact/> </app> <app_version> <app_name>llrPSP</app_name> <cmdline>-t 8</cmdline> <avg_ncpus>14</avg_ncpus> <max_ncpus>14</max_ncpus> </app_version> </app_config>
Then restart the boinc service.
P.s. Could a super mod maybe move this PrimeGrid stuff to the PrimeGrid thread?
8 and 14 ? shouldn;t that be 16 ?
14 gives you a couple of cores to do GPU work if you want.
OK< I think I have it. Does this look right ? On this box, NO GPU is running at the moment
But its only using 50% of the CPU, why ?
Edit: I changed the 8 in the config file to 16, now@93% (odd) and ETA is 1 day 3 hours
It keeps changing the ETA now down to 12 hours 5 min, one hour 32 min elapsed. Is that good ?
And I am a super mod, so once we are done, let me know what post numbers need to move, and a link to the destination thread.
@Markfw, the LLR applications run fastest on Intel CPUs if Hyperthreading was disabled in the BIOS. Almost as fast if HT is on, but only 50 % of the logical CPUs used. Slower if HT is on and more than 50 % of the logical CPUs are used.
AMD's SMT performs different from Intel's HT. Therefore whether SMT should be on or off for optimum performance of PrimeGrid LLR on Ryzen is an open question.
That's for PrimeGrid Sieve, which differs from LLR. LLR very heavily relies on AVX. While all recent AMD processors implement the AVX instructions, they have fewer hardware resources behind it than some of the Intel processors.
On my dual-socket Xeon E5-2690v4 (2x14 cores, 2x28 threads if HT on, 2x14 threads if HT off, 2.9 GHz):
HT on, 56 single-threaded jobs: ...... 3.95...4.2 days task duration .......... 13.7 tasks per day
HT off, 28 single-threaded jobs: ........ 1.9...2.0 days task duration .......... 14.4 tasks per day
HT off, 14 dual-threaded jobs: ................... 1.3 days task duration .......... 11 tasks per day
HT off, 4 seven-threaded jobs: ............ 5.5...6 hours task duration .......... 16...17.5 tasks per day
HT off, two 14-threaded jobs: ...................... 3 hours task duration .......... 15.5 tasks per day
(That's from the thread "PrimeGrid: CPU benchmarks". Take half the number of simultaneous jobs and half the number of tasks per day for a single-socket machine.)
Haswell Xeon E3-1245v3 (4 cores, 3.4 GHz):
HT on but only 50 % logical CPUs used, 2 dual-threaded jobs: .... 29 h task duration .... ~1.6 tasks per day
If the Ryzen runs 1 task in 12 hours, that's 2 tasks per day obviously, 1.2x as fast as the 4-core Haswell.
BTW, team AMD Users beat us narrowly in the PrimeGrid GCW-Sieve race in January, but they are far behind us in this PSP-LLR race. Sieve going better on AMD processors than LLR could be a reason.
[link to the challenge series]
OK, I am doing a unit in 6 hours 45 minutes. How is that ?
That's just wrong. Because my i7 is taking 10 hours to do a WU.
That is really, incredibly good! Are you overclocking at all? Because that could cause a WU to fail or be invalidated.
Well, its been validated I think by WCG running 24/7 like this for a week. It was @4.0, but it was flaky at that speed.
See my post in the primegrid thread. (and I couldn't figure out how to move posts, maybe thats an admin thing)
Edit: a 6700 has 8 threads, and this has 16, so 5 hours would make it even, and its a little more than that. Currently about 6.5 hours, but it keeps going down in estimated time. Maybe I will try my E5's and see how they do.
Be sure to either set the -t value to the real number of cores on an Intel processor, or turn off HT.