Separate names with a comma.
Discussion in 'Distributed Computing' started by Ken g6, Dec 29, 2016.
6 hours 20 minutes for this one per 67% quesstamate.
One CPU having to access memory which is located at the other CPU would not cause an error, it would merely slow things down. On the other hand, it is possible that the programmers already took precautions specifically for a multithreaded task which runs on two sockets. Here is a place were we could ask to find out.
A pro pos: Right at the end of that thread is a posting which says that multithreaded tasks are in fact properly balanced over physical cores and don't wander around (unlike several concurrent processes which are not being balanced as good and, to add insult to injury, tend to migrate from core to core on Windows). The poster presumes that this positive side-effect of running LLR multithreaded is thanks to an underlying library.
So it would seem that setting affinity is not necessary, which is good, I don't have time to babysit the machines all the time. I was able to get a monitor over to one of the dual hexacores and shut off HT, let's see how its results compare to the other dual hexacore, which runs faster clocks but is also doing GPU and is stuck with HT on. Very scientific!
They have 940...950 k PPD yesterday and today. If they continue that way, we would need more than 1.74 M PPD to catch them. We have got 1.27 M PPD today.
Spoiler: User graph
Spoiler: Team graphs
Tony, what is the config file for your 2683's, and is HT off, and whats the PPD ? I have 4 of them that could make quite the difference, and one more Ryzen. at 15k per 6 hours, that 60k ppd I think for the one Ryzen. And we need almost 500k ppd to catch Poland ?
@Markfw, in my experience it is safe to write a new app_config.xml while BOINC is already crunching, then shut the manager down with "[x] Stop running tasks when exiting the BOINC Manager" ticked, then restart the manager. It then continues the existing WUs at the very percentage where it left off, but with the new threading options applied (and if necessary running fewer or more tasks simultaneously). That is, you don't lose prior work.
There is even a way to change app_config.xml without a full restart (I described it further up this thread), but that way doesn't really buy you anything over a restart and can have the downside of misleading displays of the CPU usage.
OK< at 30 minutes it was not even at 1%, so I disabled HT, and made a app_config.xml file with 14 in the 3 places. Its back to running.
Update. Looks like 6 hours ETA at the moment. The Ryzen is NOT faster ! That could change though
Hmm, my i5-6500 is going to take 28 hours to finish one WU using all 4 cores, something is wrong. I don't see any warnings in the log. I thought Skylakes were supposed to be good at this project.
Well, the benchmark I saw said that Ryzen was barely better than Skylake and I have 4 times the cores, and probably more MHZ. Even if the mhz was even, 4 x mine is 26 hours. So I hate to say it, but it right Its think. Your dual E5660's might kick some hiney though, if configured to run like mine, all cores in one task. Thats 12 at 3 ghz with HT off, or my estimate would be not far from my Ryzen.
Well, here's the weird thing. I have an i5-4690 that says it's going to finish one in about 14 hours, and that is with only 3 cores because it's overheating on 4. So the discrepancy is what is making me wonder.
(I'm going to slap a bigger cooler on the Haswell tonight and go back to 4 cores)
Coincidentally, this first WU had been previously crunched by Tony, thus validated by Mark.
Sounds like the performance of two cores rather than four. Does task manager confirm that all four cores are near fully used? Anything else running on this host?
No, I just don't understand it. I rebooted just to see if that would help, negative:
The screenshot does not show that all the other projects ARE suspended.
Are you sure your RAM is running dual-channel?
Mine are running with HT still enabled, but I have configured BOINC to use only 50% of the processors, and the app_config helps to keep usage under 50% as well. I had originally planned to use my GPU's during this race, and wanted to save threads for them, so my app_config is a bit under performing as I have it set to run three tasks with each using 4 cores, leaving two cores for the GPU. A quick snapshots shows they all are on track for completion in 14 hours. So if I did the math right, 3 tasks every 14 hours, per box is about 77,000 ppd.
I am NOT using GPU though, since I lost my mind and fired up 9 other computers for the Numberfields race, and my electric bill needs a break. So I should revisit my app_config I suppose, and get those other two cores per box to doing something.
Oh, now that you mention it, this board has a damaged socket pin that killed a channel! I never realized it would make that much difference! Dang...
I disabled SMT for the Ryzen, since I am now getting like a little over 3 hours with NO HT on the 2683 ! Ryzen was at 6-7 hours, so I thought maybe it would help. Well, no, and now that its re-enabled, Windows won't see all 16, only 8 ! Damn windows !!!!
Oh, silly goose, just restart it again. It usually only takes 3-4 restarts for Windows 10 to see a (same model but) different video card.
I restarted many times. Did not work, reloaded bios defaults, did not work. Flashed the BIOS to what it already was, and BAM, its back.
But Tony, look at this !! My E5-2683 is doing these in 2:50 !! app_config set to 14,14,14, NO HT
Here is my app_config.XML
<app_config> <app> <name>llrPSP</name> <fraction_done_exact/> </app> <app_version> <app_name>llrPSP</app_name>
<cmdline>-t 14</cmdline> <avg_ncpus>14</avg_ncpus> <max_ncpus>14</max_ncpus> </app_version> </app_config>
And yes, this is at stock, these Xeons don't OC
The box which I have currently running -t 14 needs 3h30m right now. These are WUs with 16,300...16,500 points per WU.
Half a week ago it was less than 3h:07m as an average, but it had WUs with 10,500 and 15,700 points mixed in together with 16,300 ones. So, WUs are definitely variable.
@Markfw, here is a WU of yours with merely 12,100 points credit, completed at 21:26 UTC. Was this already with -t 14 from start to finish?
I have got DDR4-2400 cl17 registered ECC RAM, is yours faster?
Judging from the report time, the WU which I linked could indeed be the one highlighted in your screenshot.
Did you OC the RAM?
Can't OC the RAM, processor locks it to 2133max on V3 chips, 2400max on V4 chips.
The E5-2683 was done from less than 1% to finish with -t=14.
The ram is 2133@stock Corsair ram cas 15
The 2683 is on unit 2 now, and so is the Ryzen, but the 2683 will zip right past it at this rate.