Too many iGPU, not enough real GPU tasks, etc.

Message boards : GPUs : Too many iGPU, not enough real GPU tasks, etc.
Message board moderation

To post messages, you must log in.

AuthorMessage
Mike Ryan

Send message
Joined: 7 Dec 16
Posts: 4
United States
Message 74637 - Posted: 7 Dec 2016, 6:00:30 UTC

I have a couple of minor issues with BIONC (or SETI@Home?) that I'm not finding by searching so I hope I'm not repeating information that has already been posted.

Running: BIONC v. 7.6.22 (x64) client on Windows 10 Home.
i4590 CPU @ standard speeds, 4 cores, (plus HD 4600 graphics)
GTX 1060 6GB graphics card (standard speeds)
16G memory
* the only project I run is SETI@home

Issues with SETI@home tasks:
1) BOINC does not seem to run tasks that are about to expire in a "higher" priority than newly downloaded tasks. It doesn't seem to do so in a simple FIFO order either. I just had 5 iGPU tasks expire today (12/7/16 in UTC) that were sent to my computer way back on 10/24/16. Each of the iGPU tasks generally process in a little over an hour - and the computer is on 24/7. How is it possible that after literally thousands of other tasks have processed since downloading these few, they did not yet run (and then expired)?

2) I also seem to get too many CPU/iGPU tasks vs. the number given to my GTX card. At any given time my current cache might consist of the following:
90 CPU tasks of SETI@home v8. 8.00 (~2.5 - 5 hours per task, on average completes about a task/hr)
100 iGPU tasks of SETI@home v8 8.19 for opencl_intel_gpu_sah (~1 - 1.25 hours per task)
100 GPU tasks of SETI@home v8 8.19 for opencl_nvidia_SoG (~5 - 12 minutes per task)

My preferences are set to (a) crunch AstroPulse tasks, but allow others if no work available (there rarely is), and (b) store "at least 5 days" with up to 10 additional. Using all cores at all times.

So for the regular CPU cores and the iGPU, I have ~3-4 days of work queued, but the 1060 even with all "worst case" tasks will eat through 120+ of them in less than a day. In order to be on par with the CPU and iGPU tasks, I'd think there should be closer to 500 Nvidia tasks queued up. Is the "100" some sort of hard limit imposed by the SETI folks perhaps?

Thanks for any insights.
ID: 74637 · Report as offensive
robsmith
Volunteer tester
Help desk expert

Send message
Joined: 25 May 09
Posts: 1283
United Kingdom
Message 74638 - Posted: 7 Dec 2016, 8:11:05 UTC

This apparent imbalance is due to the way SETI@Home sets its limits combined with your cache settings.
S@H has a "hard" 100 task limit for the CPU, plus an additional 100 tasks per GPU.

The problem comes from your cache settings - bitter experience says that by having a LARGE additional days setting causes problems with getting things right. You would do better (and I know this sounds counter-intuitive) to reduce the "extra days" to less than one, this would mean you would get the maximum of five days work for the CPU, and 100 tasks for the 1060, and probably slightly less than 100 tasks for the iGPU (depending on its calculated performance - which can be wildly out)

Attempting to "micro manage" the order in which tasks are run will almost always end up in tasks timing out.
Further - your iGPU is not as fast as you think it to be - one of the "valid" results I could find took 11,000 seconds to run, which is about 3.5 hours, the others were somewhat faster (this is in line with what I saw when running an i7 - the iGPU wasn't much faster than a CPU core, but generated a lot of heat in getting there causing the whole processor to slow down until it had cooled enough to run flat out...)

(btw - this has been extensively discussed on the SETI boards where this, and similar questions, comes up every few weeks. And S@H has not indicated that there are any plans to change the limits)
ID: 74638 · Report as offensive
Mike Ryan

Send message
Joined: 7 Dec 16
Posts: 4
United States
Message 74676 - Posted: 8 Dec 2016, 6:23:23 UTC - in response to Message 74638.  

Thanks for the information. Like I mentioned, it looked like "100" was some kind of hard limit. I'll try adjusting the upto X extra days to a fraction of a day and see if that makes any difference.

And yes, the iGPU varies greatly in how long it takes. But I'd guess that in at least 80% of the time it kicks out a work unit in about 1.25 hours. The CPU tasks seem to take around 50% longer to process while the iGPU is also running. (I doubt a heat issue as I track that, but more of a very busy cache issue.) I'm currently experimenting over a several week period to see if I'm better off crunching with or without the iGPU.
ID: 74676 · Report as offensive

Message boards : GPUs : Too many iGPU, not enough real GPU tasks, etc.

Copyright © 2024 University of California.
Permission is granted to copy, distribute and/or modify this document under the terms of the GNU Free Documentation License, Version 1.2 or any later version published by the Free Software Foundation.