Unbalanced load on HD7950s why?
log in

Advanced search

Message boards : Number crunching : Unbalanced load on HD7950s why?

Author Message
Profile BeemerBiker
Avatar
Send message
Joined: 18 Nov 08
Posts: 94
Credit: 566,616,443
RAC: 2,920,108

Message 67039 - Posted: 8 Feb 2018, 15:16:42 UTC
Last modified: 8 Feb 2018, 15:22:51 UTC

I have a pair of HD7950s. Unfortunately (I now know to check for engineering samples) the one I got from eBay has only 1536 shaders while the "normal" one has the full 1792. Both have full 3gb of memory.

I set up app_config to run 5 instance of MilkyWay on each card. Unaccountably, GPU0 has 5 instances, but GPU1 has 4. I have plenty of memory available , 12gb and there are extra cores not being used.

GPU0 was the lamer that I got on ebay and it seems to run hot so I swapped cards in their pcie slots and rebooted. I noticed the 5 tasks (not the same ones) are still on GPU0 and that is now the "normal" card. There are 4 on GPU1 (not the same 4) which is the lamer.

Tasks come and go so quickly that is is hard to snag an image. However, as long as I have watched different WUs came and go, I never see more than 4 on the "GPU1" card even though the tasks types change. FWIW, the outer PCIe slot is x8 in an X16 while the one closest to the CPU is full X16. There is no crossfire enabled.

Why is that? Just another puzzle it seems.

mmonnin
Send message
Joined: 2 Oct 16
Posts: 99
Credit: 79,301,187
RAC: 63

Message 67041 - Posted: 9 Feb 2018, 13:30:30 UTC - in response to Message 67039.

Is this your PC or a screenshot from months ago? You have no tasks.

Some of them are 2-4 core count machines. What happens if you get some tasks and change it to 0.1 CPUs in app_config.

Profile BeemerBiker
Avatar
Send message
Joined: 18 Nov 08
Posts: 94
Credit: 566,616,443
RAC: 2,920,108

Message 67043 - Posted: 9 Feb 2018, 16:43:40 UTC - in response to Message 67041.

I am still stuck in grcpool until I get enough GRC coins to go back to BAM! and have my credit and tasks show up like everyone else here.

I think the problem is the boards are not exactly identical. I don't have this problem on a Ubuntu system with pair of identical 1050i. It is just strange, some artifact of slightly different cards I am guessing. I suspect the boinc developers have bigger problems than to look into this. Was just wondering if anyone else had seen this. I am losing some efficiency by not being able to run a 5th task on the second board.

mmonnin
Send message
Joined: 2 Oct 16
Posts: 99
Credit: 79,301,187
RAC: 63

Message 67044 - Posted: 9 Feb 2018, 19:16:23 UTC

I've ran AMD and NV in the same machine and they run the correct # of tasks. If anything its your setup.

Profile BeemerBiker
Avatar
Send message
Joined: 18 Nov 08
Posts: 94
Credit: 566,616,443
RAC: 2,920,108

Message 67321 - Posted: 10 Apr 2018, 15:31:20 UTC - in response to Message 67044.
Last modified: 10 Apr 2018, 15:43:42 UTC

This is an old post but I found a solution to my unbalanced load. I used .19 instead of .20 for the gpu work distribution. .20 should have given exactly 5 units but there must be some rounding error or some math computation in OpenCL that caused 4 units to show up on one board and 5 on the other. Using 0.19 allowed 5 work units on each board. Each board was capable of running 8 units as I have experimented with 0.125 distribution.

I also ran the following test for what it is worth. This was on S9000 which similar to HD7950 and has the same driver, OpenCL 1.2 and NOT 2.0

I set distribution to .125 and only allowed one board to be used. There were no CPU or GPU tasks running. I then unsuspended one task and made a note of the time to complete. I then unsuspended 2, then, 3, etc. When I got to 6 concurrent tasks the (averaged) time to complete became more for each thread instead of less. So, at least for my setup, 5 threads (Milkyway units) gave the best performance on my S9000 which, except for the extra memory, is identical to hD7950.


Post to thread

Message boards : Number crunching : Unbalanced load on HD7950s why?


Main page · Your account · Message boards


Copyright © 2018 AstroInformatics Group