Message boards :
Number crunching :
Ideal number of concurrent tasks.
Message board moderation
Author | Message |
---|---|
Send message Joined: 5 Mar 14 Posts: 24 Credit: 500,964,006 RAC: 0 |
On the higher FP64 cards like the tahitis, and radeon vii, some Titans, it seems better to run more than one task at once, but how many is optimal? |
Send message Joined: 9 Dec 11 Posts: 38 Credit: 1,497,896,956 RAC: 0 |
It really depends on *which* high end card you're actually talking about. In my dealings 'round the DC circles it's been said that Nvidia cards see little gain in running multiple concurrent WUs. The AMD cards are a different story. I can only personally speak to the Vega64, RX580, and Pro VII. I run both the RX580s and Vega64s with 2 concurrent WUs each. My Pro VIIs run 4 concurrent tasks. I think i may be able to squeeze more out of the Vegas and VIIs but I just haven't gotten around to it. I believe part of the equation is the High Bandwidth Memory those cards use. The best way to figure it out is either talk to somebody with your specific card(s) or take notes on runtimes for single WU operation, then set it to double and see if your WU/time rate improves. |
Send message Joined: 30 Dec 14 Posts: 34 Credit: 909,988,687 RAC: 17 |
I have four Radeon HD 7990 cards (2 GPU each card) running 5 WU/GPU. My CPU is a 4 core Intel i5 7600 (no hyperthreading), so I have 0.20 GPU and 0.01 CPU in my app_config.xml file. |
Send message Joined: 30 Dec 14 Posts: 34 Credit: 909,988,687 RAC: 17 |
duplicate |
Send message Joined: 5 Mar 14 Posts: 24 Credit: 500,964,006 RAC: 0 |
I have four Radeon HD 7990 cards (2 GPU each card) running 5 WU/GPU. My CPU is a 4 core Intel i5 7600 (no hyperthreading), so I have 0.20 GPU and 0.01 CPU in my app_config.xml file. I'm running two 7970s in one machine with a 3770k @ 3x WU each, 6 total. I can run 5-6 threads of universe and not slow down the MW tasks, at 7 threads it holds for cpu every once in a while. MW uses very little cpu, way less than the default suggests. I've dabbled with anywhere from 3-5 per 7970 over the years. my reasoning is mostly I wanted to reduce time in between WUs because at 1x they finish so fast that the couple seconds between wus becomes meaningful credit loss, imo. seems like i get about 440k per day on the ghz one and 420k on the reference one @ 3 each. ill go to 5 and see if i squeeze any more out. not sure why but after i set my app config to 0.20 i only got a total of 9 wus total, i switched to 0.19 and got 10 wus total, hopefully 5 each. |
Send message Joined: 9 Dec 11 Posts: 38 Credit: 1,497,896,956 RAC: 0 |
Ya, I noticed something similar. You post got me to get off my duff and set mine from .25 to .20. Until I read your reply I thought it was odd that some of my WUs were still completing in the time they were at .25, yet others were taking a tad longer. |
Send message Joined: 18 Jul 10 Posts: 76 Credit: 635,998,708 RAC: 0 |
I monitor the GPU processor usage. GPU memory usage has never been an issue for me. There are various ways to accomplish this based on your operating system and video card. I figure if the GPU processor usage is substantially over 50% with one task, it probably doesn't make sense to run more than one task concurrently. If the GPU processor usage is less than 50%, running additional tasks would increase my throughput. |
Send message Joined: 5 Mar 14 Posts: 24 Credit: 500,964,006 RAC: 0 |
Ive tested 3x 4x and 5x and seem to get pretty consistent credit. i could go down to two but im confident i wont get better credit. with higher concurrency i collect more WUprop hours. |
Send message Joined: 13 Feb 11 Posts: 31 Credit: 1,403,524,537 RAC: 0 |
I'm running 8x on my Radeon VII but the real problem.... is the "Project" is not willing to keep it fed with work 24/7 ! I should sell it to some Bitcoin nut-job and scalp the hell out of them.... Grumpy old man, dunx |
Send message Joined: 21 Feb 18 Posts: 3 Credit: 1,806,236 RAC: 0 |
I'm running 8x on my Radeon VII but the real problem.... You probably know more about this than I do, but I've heard there is a way to run multiple clients on the one PC so that you don't hit limits imposed by the project. Or are you saying the project is running out of work? You could also crunch additional GPU project if you wanted. |
Send message Joined: 24 Jan 11 Posts: 708 Credit: 544,757,713 RAC: 71,923 |
there is a way to run multiple clients on the one PC so that you don't hit limits imposed by the project This is correct. Just run multiple clients and double, triple or quadruple the number of tasks you get per day because you are telling the scheduler you have 2X. 3X or 4X the number of gpus so the scheduler knows it has to feed more than your one R7. Or spoof the number of cards in your coproc_info.xml file to tell the client you more than one R7. Here is somebody's recipe for creating a multi-client host. https://www.overclock.net/threads/guide-setting-up-multiple-boinc-instances.1628924/ |
Send message Joined: 8 May 09 Posts: 3325 Credit: 521,126,154 RAC: 39,574 |
I monitor the GPU processor usage. GPU memory usage has never been an issue for me. There are various ways to accomplish this based on your operating system and video card. Generally speaking, it really depends on memory after one unit at a time, but if it under 85 to 90% usage for one task trying running another task, 50% is leaving ALOT of crunching ability on the table. Now crunching more than one unit at a time also pushes the gpu towards it's upper limits and doing so can lower the life span of the gpu, miners do that all the time which is often why gpu's on the aftermarket that were used for mining don't do nearly as well as we had hoped they would. Like all those 30XX gpu's that were gobbled up before we crunchers had a chance to get them that will be coming back onto the market once there is something better or the bitcoin price goes too low, be careful of what you are buying. |
Send message Joined: 13 Feb 11 Posts: 31 Credit: 1,403,524,537 RAC: 0 |
Thanks, for the info guys.... I was of the opinion that multiple clients wouldn't be allowed. The project sends 900 tasks which process in about 10 seconds each, then it sits for 15 minutes to pass, before sending more. At my desk I can poke it into action quite easily, but all day whilst at work it's on autopilot. Must investigate further ! Regards, Dunx |
©2024 Astroinformatics Group