Binar radio Pulsar search. taking a while to run
28 Jan 2015 18:25:09 UTC
Topic 197951
(moderation:
I'm, having a few issues with these. They load saying expected run time of 00:46:46. After about 5 hours it seems that there is about 16 seconds remaining. (00:00:16), after a further 2 hours there is 1 second remaining and at 8 hours there is zero seconds remaining. It then stays at zero seconds remaining with progress of "100.00%" for a long time. Current elapse is 10:50:10 and counting. I've been running these for about 3 weeks now and they seem to "end" when I'm not looking (i.e. in the middle of the night) so I have no idea how long it goes on for.
Also between a third and half of them exit with "computation error".
I'm running W 8.1 on quad core.
Any ideas anyone?
Binar radio Pulsar search. taking a while to run
)
What version of the graphics driver are you using?
Most of your errors occur right at the start of the tasks and is probably driver related. Read this post by Claggy for links to download a know good driver from Intel.
Apart from that it's a known "feature" that the Binary radio pulsar tasks, or BRP4 for short, when running on an Intel GPU needs a full CPU core to support it or else the performance suffers greatly.
The easiest way to achieve this is to go to your account and then edit the computing preferences and change "On multiprocessors, use at most XX% of the processors" to something less than 100%. In your case with 4 CPU cores set it to 75% to leave one core free to support the GPU app.
Or if you are using local preferences set through Boinc manager open Boinc in advanced view and in the "Tools" menu choose "Computing preferences", then on the "Processor usage" tab change the same value near the bottom. Take note that local preferences always override web based preferences.
As to the expected runtime that's just a guess and if your running more than one app per project it will have a hard time to adjust to reality. Better look at elapsed time and percentage done to give a hint on how long the task will take. I complete these tasks on the same type of Intel GPU (HD4000) in about 13-15 min per task so that's your target.
RE: I'm, having a few
)
Boinc 7.2.38 and later will Estimate progress up until the app checkpoints for the first time:
client: if app doesn't report fraction done, estimate it
and it'll also:
client: if app doesn't report fraction done, estimate fraction done in a way that converges to but never reaches 100%.
Meaning the longer it runs, and never reach the first checkpoint (and so report it's real progress), the slower it'll seem to progress,
I think that's what's happening in your case, the app is stalled and not making progress and never reaching the first checkpoint,
So Boinc continues and continues to Estimate the apps progress, try freeing a CPU core, and/or try a different driver.
Claggy
Thanks for that. I set it to
)
Thanks for that. I set it to 75% cpu & 75% of cores and it now goes through at about 11 hours. I also tried some things. I put a hold on all new jobs (this is my laptop, so no mega power, I just run Einstein and malaria control). It got to the point when only one job (the BRPS) was running. Other than that I am using it for just a bit of work on Word & EMails. It still took 9 hours to complete. The CPU was running at about 2% (two percent) busy, with each core running at near to nothing. No major disk, memory or network. Elapse time was running normally and the "time remaining" was going down at about 1 second per 20-30 seconds elapse, and the nearer it got to zero the slower it got. At remaining = 10 seconds it was taking over 50-60 seconds to reduce the remaining time by a single second.
It obviously does not like my machine or the way it is set up.
The thing is that it seems to clog the computer and if it goes to sleep (being a laptop that can often happen) it frequently restarts the BRPS job or as previously stated, aborts it (6+ hrs of processing down the drain).
As I can't seem to get anything other than BRPS from Eistein at the moment, I think I'll just have it as "no new tasks" for a while.
Thanks for your help everyone.
Yes, it is one of the
)
Yes, it is one of the oddities of running the Einstein Intel GPU application that reported CPU consumption is tiny, and yet all other indications show very strong influence of CPU resource conflict as though the "real" CPU use in some sense were much higher. Possibly this takes the form of competition for shared memory resources or data connections, use of which by the GPU is not "charged" as CPU time, but which nevertheless competes with other CPU use.
This goes both ways--heavy CPU use greatly slows the progress of the Intel GPU job--sometimes gravely, and the GPU job combined with full CPU use has slowed interactive performance to an annoying degree.
My host with a modern Intel GPU plus capable nVidia GPU gave me so little net productivity increase and so much puzzlement that I disabled use of it. That said, it is quite likely that I failed to find the "magic settings" which would have reaped decent additional output from it.
At the time I was running, I don't think I had a big invalid result problem, but I am pretty sure all the time I was running I had a constraint on total CPU jobs, possibly a pretty severe one. Quite likely I was also not running the same Intel driver with the same Einstein application version as you report.
As with your host, mine had a two-core hyperthreaded CPU. My current doctrine on hyperthreaded CPUs on hosts with graphics processors running Einstein is that I load them with no more than the number of CPU jobs that they have physical cores, and use Process Lasso-applied CPU affinity to restrain those explicit CPU jobs to just one side of each core. That stops the "unlucky assignment" problem of two CPU jobs ganging up on one physical core while leaving another idle, and tends to give low latency to the GPU-servicing CPU tasks, which get free call to either side of all cores. It nevertheless consumes the great majority of the available CPU capacity, while keeping the (much more efficient) GPU jobs running about as fast as I can get them to go.
Well I let the one last unit
)
Well I let the one last unit work. Except that it re-started- twice (having done 9:45:00+ elapse on each), so I let it go,(third time lucky?) and having got 9:40:00 elapse and 00:00:00 outstanding,100% complete and nothing else using CPU I thought I was there. 3 hours later, with still 100% complete, 00:00:00 remaining it had not completed and then it said "computation error".
So I said "enough". My Einstein processing will have to wait until units do not require my GPU. I have set Einstein to "no new task". I'll look again in a couple of months.
Thanks everyone for your input.
Then I would suggest that you
)
Then I would suggest that you edit your Einstein@home preferences and under "Run only the selected applications" opt out of "Binary Radio Pulsar Search (Arecibo)", these are the tasks sent to your Intel GPU. Your computer is capable of completing CPU only tasks as shown by your tasks list filtered on FGRP tasks so you can still make a valuable contribution.
You've been with the project for several years and have made a big contribution, we would miss you if you left!