BRPS (Arecibo,GPU) WUs running on CPU!!!

Cruncher-American
Cruncher-American
Joined: 24 Mar 05
Posts: 71
Credit: 5,269,847,980
RAC: 4,375,390
Topic 227465

I have my cruncher Ralph set up to run ONLY Einstein GPU WUs. I  had never received any Binary Radio Pulsar Search (Arecibo, GPU) (1.33) WUs until today, and BOINC tried to run them on my CPU, NOT GPU.

 

What?

Harri Liljeroos
Harri Liljeroos
Joined: 10 Dec 05
Posts: 4,189
Credit: 3,109,663,823
RAC: 1,817,124

These GPU tasks are currently

These GPU tasks are currently only for Intel GPUs. But maybe we will have an application for Nvidia/AMD also in the future, see the Technical News forum for updates.

Ian&Steve C.
Ian&Steve C.
Joined: 19 Jan 20
Posts: 3,902
Credit: 43,646,762,643
RAC: 63,147,246

there is no functional

there is no functional difference between the BRP4 and BRP4G tasks, and they can use the same application to crunch both types.

 

the BRP4G tasks are just bigger/longer than the BRP4 tasks and are awarded proportionally more credit, that's it. they are intended for GPUs, but can run on a CPU just the same. it'll just take a long time to complete. if you check the applications list, you will see several CPU applications in the BRP4G "category". it might have been a mistake, but maybe not.

 

The project is ramping up to change some things regarding BRP4/BRP4G/BRP7, so things will probably be in flux as they fine tune their configs. Bernd has mentioned that BRP4 will be reserved for slow ARM CPUs, BRP4G will go to faster x86 CPUs AND intel GPUs, and faster GPUs will get a new project called BRP7.

 

PS: It will be better to refer to your systems by their host ID or simply linking to the host in question. While YOU can see hostnames listed on your hosts' page, we have no idea which system is supposed to be "Ralph", no one else can see that. we can only see the list of hosts and their stats/tasks.

_________________________________________________________________________

Cruncher-American
Cruncher-American
Joined: 24 Mar 05
Posts: 71
Credit: 5,269,847,980
RAC: 4,375,390

Thanks; I will try to

Thanks; I will try to remember that going forward.

 

And thanks for the info.

Cruncher-American
Cruncher-American
Joined: 24 Mar 05
Posts: 71
Credit: 5,269,847,980
RAC: 4,375,390

That should not have affected

That should not have affected me - I have dual hybrid GTX 1080tis on that machine (12881834).

Though it IS a Ryzen (3950x) - might that have confused BOINC? Did I detect a bug?

Harm
Harm
Joined: 24 Aug 05
Posts: 6
Credit: 16,606,027
RAC: 0

I got CPU jobs also (not

I got CPU jobs also (not Intel GPU) for Binary Radio Pulsar Search (Arecibo) while only having selected GPU jobs, so I aborted them and unselected Binary Radio Pulsar Search (Arecibo, GPU) also, please do not send jobs that were not selected, I want to operate (fast) GPU jobs only for limited hours, I will have to abort the unselected ones.

Ian&Steve C.
Ian&Steve C.
Joined: 19 Jan 20
Posts: 3,902
Credit: 43,646,762,643
RAC: 63,147,246

Harm wrote: I got CPU jobs

Harm wrote:

I got CPU jobs also (not Intel GPU) for Binary Radio Pulsar Search (Arecibo) while only having selected GPU jobs, so I aborted them and unselected Binary Radio Pulsar Search (Arecibo, GPU) also, please do not send jobs that were not selected, I want to operate (fast) GPU jobs only for limited hours, I will have to abort the unselected ones.

there is a selection in the project preferences for “allow non-preferred applications when no work available” or something like this. Make sure this is not selected in your preferences. 
 

also make sure you unselect “use CPU” in your preferences. 

_________________________________________________________________________

Ronald McNichol
Ronald McNichol
Joined: 28 Feb 22
Posts: 27
Credit: 99,853,798
RAC: 0

I find that the BRP4 is

I find that the BRP4 is especially greedy! I have my systems set for a maximum of 0.5 days of jobs, and the moment I selected the BRP4 (now called Large rather than GPU), it loaded up >179 8-hour jobs that don't seem to run on my Intel GPU (actually APU) even though the Gamma Ray GPU job does run on it. It does run on as many CPUs as I allow to be used, including all 12 threads (six cores and six ersatz cores). I now have it restricted to 5 cores, and _usually_ BRP4 (large) graciously allows one core to be used with my NVidia discrete GPU for the Gamma Ray study.

I think they should make a BRP4 that ONLY runs on the GPU, and I would re-allow my Intel one to work. As it is now, I disallowed BRP4, but it will be a month before it finishes the work it has already queued up! I looked this morning, and the queue was down to 171, running on 4 cores. I assume that running on more cores was slowing things down too much, probably due to shared memory and cache misses.

Processor    Intel(R) Core(TM) i7-8750H CPU @ 2.20GHz, 2208 Mhz, 6 Core(s), 12 Logical Processor(s)
Not a barn burner, but better than no computer! ;)

 

 

mikey
mikey
Joined: 22 Jan 05
Posts: 12,544
Credit: 1,838,590,455
RAC: 3,606

Ronald McNichol wrote: I

Ronald McNichol wrote:

I find that the BRP4 is especially greedy! I have my systems set for a maximum of 0.5 days of jobs, and the moment I selected the BRP4 (now called Large rather than GPU), it loaded up >179 8-hour jobs that don't seem to run on my Intel GPU (actually APU) even though the Gamma Ray GPU job does run on it. It does run on as many CPUs as I allow to be used, including all 12 threads (six cores and six ersatz cores). I now have it restricted to 5 cores, and _usually_ BRP4 (large) graciously allows one core to be used with my NVidia discrete GPU for the Gamma Ray study.

I think they should make a BRP4 that ONLY runs on the GPU, and I would re-allow my Intel one to work. As it is now, I disallowed BRP4, but it will be a month before it finishes the work it has already queued up! I looked this morning, and the queue was down to 171, running on 4 cores. I assume that running on more cores was slowing things down too much, probably due to shared memory and cache misses.

If you hilite one of the running tasks in Boinc Manager and then go to the left and choose Properties it will pop up a box telling you how much memory each task is using both min and max. You could also be taxing the memory that way if ie you have 16gb of ram and each task takes 2gb and you are trying to run 10 of them at once.

Quote:
Processor    Intel(R) Core(TM) i7-8750H CPU @ 2.20GHz, 2208 Mhz, 6 Core(s), 12 Logical Processor(s)
Not a barn burner, but better than no computer! ;)

I have several pc's that are slower than that one that crunch just fine, for me it's not the speed it's the number of cpu cores I can have crunching. I can't afford the latest 4.5 or 5ghz 8 core cruncher but I can afford to get 3 or 4 12 core crunchers running at 2.2ghz.

Ronald McNichol
Ronald McNichol
Joined: 28 Feb 22
Posts: 27
Credit: 99,853,798
RAC: 0

Since this message yesterday,

Thanks for the tip about properties. My concern wasn't the amount of memory taken up, but rather cache misses making the core have to wait its turn to access physical memory, or on L3 cache hits having to wait their turn. Only 30% of my 32GB of memory is being utilized running with 5 cores of BOINC jobs plus the also pretty greedy Chrome. Resource Monitor has 9 GB in use, .5 GB modified, 6 GB Standby (whatever that is), and 16GB free. Absolutely no paging or swapping occurs even when I allow 12 jobs to run. Only when faulting in new jobs.

As a test, I just turned my computing prefs up to 100%, and I am still only using 35% of memory. Now running 13 BOINC jobs out of a possible 12! LoL 12 BRP4s and one GPU job that was already running. The only effect is the runtimes being extended. I am pretty sure physical memory access is the bottleneck. ...and perhaps L3 cache. I don't have the instrumentation software to see what is going on at that level of detail.

Each job with 8 running takes about 2.5 times as long to run as they do with 4. Twice as many jobs in 2.5 times the time. Shoveling excrement against the tide! ;) 4 jobs seem to be a more productive number (plus another for the GPU).

Since this message yesterday, the # of BRP4 jobs in the queue is down to 163, and all of them have a deadline of May 8th! Funny, but I think there will still be over 100 in the queue by then! All have been in the queue since May 1. After that, as stated above, I turned BRP4 off and am just processing the ones already in the queue. 4 pages of them on a 1920 x 1080 monitor! (External monitor, not my laptop's screen).

P.S. I have my leading-edge AMD/AMD crunching*, and 3 Raspberry Pi 4s as well. 1 running 2 32-bit jobs, 1 running 4 32-bit jobs, and one running 4 64-bit jobs. The one running only 2 jobs is my file server I added to the mix a week or so ago.

* When I am not flying in FS2020, or playing Alyx in VR. ;)
I have BOINC set to go to sleep when it sees FS2020 running.

 

 

Comment viewing options

Select your preferred way to display the comments and click "Save settings" to activate your changes.