Message boards :
Number crunching :
Run Multiple WU's on Your GPU
Message board moderation
Previous · 1 . . . 4 · 5 · 6 · 7 · 8 · 9 · 10 . . . 11 · Next
Author | Message |
---|---|
Send message Joined: 5 Jul 11 Posts: 990 Credit: 376,143,149 RAC: 0 |
My Radeon HD 7970 seems to run at 98% GPU usage with just one task. Is multiple units no longer required due to better coding at the Milkyway end? Or are some cards different to others? I deliberately got a card that does double precision very fast compared to the price. As for the number of workunits - I remember when there was a limit of 80 at once (per machine or card I forget which). Now I seem to get 600 at a time, but it waits until they're all done before I can get another load. Although if I pause the GPU when I'm playing a game, then I seem to get a topup. How is the download limit calculated? |
Send message Joined: 23 Feb 09 Posts: 28 Credit: 10,775,220 RAC: 0 |
I'm running 5 wu's at once on my three old NVIDIA GTX Titan SC, bought on March 2013 <app_config> <app> <name>milkyway</name> <max_concurrent>50</max_concurrent> <gpu_versions> <cpu_usage>0.1</cpu_usage> <gpu_usage>0.2</gpu_usage> </gpu_versions> </app> </app_config> --------------------------------------------------------- 2x Xeon E5405 a 2400MHz 3x GTX Titan SC CPU:99% GPU1:93% GPU2:99% GPU3:95% GPU1 Temp:71 grd Celsius GPU2 Temp:71 grd Celsius GPU3 Temp:59 grd Celsius CPU1 Temp 36 grd Celsius CPU2 Temp 48 grd Celsius the machine is air coold |
Send message Joined: 5 Jul 11 Posts: 990 Credit: 376,143,149 RAC: 0 |
Are those Titan XP cards? Sorry, didn't notice you said SC. I figured XP from the speed you were getting. Can't find anywhere to compare the two. If it's similar to the XP, you should be running a single precision project like Einstein on it. It's great for that, but rubbish at double precision for Milkyway. My list of cards and how good they are: https://www.dropbox.com/s/fthl41hovchfye2/gpu.xlsx?dl=0 |
Send message Joined: 8 May 09 Posts: 3339 Credit: 524,010,781 RAC: 0 |
Thanks for the optimizing tips for the Radeon VII. MAYBE...the thing is the MilkyWay Server won't send any tasks until after youhave a 5 to 10 minute break in communications from them, hence the running out of wu's for everyone. Now if you could set it so MilkyWay does NOT report the completed wu's as completed until after that time out time then yes you could get wu's on a regular recurring basis.There is a person who provided a custom version of Boinc in another thread and a timeout script who says he's getting no breaks from crunching, I don't know if there is a way to write a script preventing MilkyWay from reporting completed tasks while still letting other projects report their completed wu's, or if it even needs too as often as it does. I know alot of us a long time added <report_results_immediately>1</report_results_immediately> to the cc_config file so we could get the rac more stabilized but here at MilkyWay it seems to report them pretty quickly even without it. |
Send message Joined: 5 Jul 11 Posts: 990 Credit: 376,143,149 RAC: 0 |
The problem would also be solved with more work units. The reason our clients are continually pestering the MW server is they want say the default of 1 day of buffer. Every other project is happy to provide 1 or 2 WEEKS of work, why isn't Milkyway? |
Send message Joined: 8 May 09 Posts: 3339 Credit: 524,010,781 RAC: 0 |
The problem would also be solved with more work units. The reason our clients are continually pestering the MW server is they want say the default of 1 day of buffer. Every other project is happy to provide 1 or 2 WEEKS of work, why isn't Milkyway? Not EVERY other Project but I don't know the answer. Some Projects base the next set if of workunits on the last set of workunits so don't provide too many, others just go thru a list until it's done so can provide lots of workunits. I believe MilkyWay gets their raw data from somewhere else but don't know how much or how often. |
Send message Joined: 5 Jul 11 Posts: 990 Credit: 376,143,149 RAC: 0 |
What do you mean by "base the next set of workunits on the last set of workunits"? Milkyway seems to be fixed at 300 per GPU, 900 per client. I've always dutifully returned all mine completed properly in a few hours, despite the time limit being a week. I should get more! I don't have this problem with Einstein which gives me as many as I ask for. If Milkyway didn't have a tiny limit, my client wouldn't be trying to get more every 1.5 minutes, so I wouldn't be using up server resources, and I wouldn't be triggering limitations on sending more units out. P.S. are those 17 computers all yours or are they at your work? 17 computers is a lot of electricity! I used to have 500 running at my work, unfortunately no GPUs. and unfortunately I don't work there any more. |
Send message Joined: 24 Jan 11 Posts: 715 Credit: 556,192,179 RAC: 55,236 |
What do you mean by "base the next set of workunits on the last set of workunits"? Some projects base the next set of tasks that the previous set results produced new parameter sets or hints about how to format the next work. Gpugrid is a project that works this way for example. |
Send message Joined: 8 May 09 Posts: 3339 Credit: 524,010,781 RAC: 0 |
MilkyWay, under the last Admin, had a problem with people storing 10 days of wu's and then them not getting crunched and having to all be recycled so they reduced the total number each device could get, this past year they raised it to where it is now. They said they discussed it in house and were happy with it but knew some people would not be, and they understood that but the decision stands. And yes all 17 of those pc's are at my house and yes the electric bill is high especially when it's gets into the 100's here at the beach on the East Coast of the US. |
Send message Joined: 5 Jul 11 Posts: 990 Credit: 376,143,149 RAC: 0 |
That's odd because Boinc shouldn't be downloading more than can be crunched on time. I don't see it happening with other projects. I doubt there were that many messing about. If it was me, I'd have just clamped down their accounts only, reduced them to lower quantities at once. |
Send message Joined: 2 Jul 09 Posts: 27 Credit: 253,069,838 RAC: 0 |
I'm attempting to run 2 WUs each on twin NVidia 1070s only, no CPU tasks, on an Intel 7700K under the current Win 10 Pro with the following app_config.xml <app_config> <app> <name>milkyway</name> <max_concurrent>4</max_concurrent> <gpu_versions> <gpu_usage>.5</gpu_usage> <cpu_usage>1</cpu_usage> </gpu_versions> </app> </app_config> I'm keep getting the following line in the event log 02/01/2020 17:49:59 | Milkyway@Home | Your settings do not allow fetching tasks for CPU. To fix this, you can change Project Preferences on the project's web site. Currently, it's defaulting to (running 0.96 CPUs + 1 NVIDIA device 0)) and device 1. What to I need to rectify this? I don't buy computers, I build them! |
Send message Joined: 5 Jul 11 Posts: 990 Credit: 376,143,149 RAC: 0 |
I'm attempting to run 2 WUs each on twin NVidia 1070s only, no CPU tasks, on an Intel 7700K under the current Win 10 Pro with the following app_config.xml That should work, I have a very similar setup - two GPUs running GPU tasks only, 2 each (I forbid CPU tasks on the Milkyway website). My app config is almost the same as yours (are you out of CPU cores? Are you running other projects on CPU? If so reduce the cpu_usage in app_config), I use this: <app_config> <app> <name>milkyway</name> <gpu_versions> <gpu_usage>0.5</gpu_usage> <cpu_usage>.45</cpu_usage> </gpu_versions> </app> </app_config> Your event log you can ignore, it's just pestering you to take CPU tasks, which you don't want anyway. The "0.96 CPUs + 1 NVIDIA device 0" could be out of date, it doesn't always change what it displays straight away until it gets new tasks. If it's only running 2 tasks on your 2 cards, instead of 4 tasks in total, you need to tell it to read the config file. If you're using Boinctasks, click the extra menu, then read config files. If you're using plain Boinc, click the options menu, then read config files. It should take effect immediately and run 4 at once, if not, restart. |
Send message Joined: 2 Jul 09 Posts: 27 Credit: 253,069,838 RAC: 0 |
I'm running basically the same setup in SETI, with 4 GPU tasks and 1 CPU task. The manager is set at 70% CPU. I'm also running current Lunatics, and the manager is 7.6.13 (x64), but I don't think that either should be relevant to my situation. When SETI has no GPU tasks available, I had hoped to run Milkyway the same way. I don't buy computers, I build them! |
Send message Joined: 5 Jul 11 Posts: 990 Credit: 376,143,149 RAC: 0 |
I'm running basically the same setup in SETI, with 4 GPU tasks and 1 CPU task. The manager is set at 70% CPU. I'm also running current Lunatics, and the manager is 7.6.13 (x64), but I don't think that either should be relevant to my situation. Your "(running 0.96 CPUs + 1 NVIDIA device 0)) and device 1." does not equate with your app config of "<gpu_usage>.5</gpu_usage> <cpu_usage>1</cpu_usage>" It has not taken into account what you told it. Restart the computer, then you're sure it's read the config you gave it. Then if it still says 0.96 and 1 instead of 1 and 0.5, then abort all tasks and let it get new ones. If it still doesn't work, you might have made a slight error in the config file. I've done that before, the slightest mistake in brackets etc and it ignores the whole lot. I have done the same as you with Milkyway and Einstein as a backup (as Milkyway has 10 minute gaps between batches of tasks it will hand out). |
Send message Joined: 2 Jul 09 Posts: 27 Credit: 253,069,838 RAC: 0 |
It seems that the app_config.xml is being completely ignored. There is no line in the event log that states that the manager recognized it. I don't buy computers, I build them! |
Send message Joined: 2 Jul 09 Posts: 27 Credit: 253,069,838 RAC: 0 |
I'm running basically the same setup in SETI, with 4 GPU tasks and 1 CPU task. The manager is set at 70% CPU. I'm also running current Lunatics, and the manager is 7.6.13 (x64), but I don't think that either should be relevant to my situation. Problem resolved. I keep certain files in a separate folder for use at a later time in case something catastrophic happens. It seems that when I copied this file into my project folder, I didn't pay attention to the full file name. Adjusted the name and it is working as expected -- A man's mind grows dim in his later years. I don't buy computers, I build them! |
Send message Joined: 7 Mar 20 Posts: 22 Credit: 106,079,902 RAC: 11,687 |
Just curious to see if anyone has really studied whether there's an appreciable gain in running >1 WU/GPU? New here at MW@H, but in previous tests at other projects I never saw any significant gain. Regards, Jim ... |
Send message Joined: 8 May 09 Posts: 3339 Credit: 524,010,781 RAC: 0 |
Hi Mikey, But it is true at several Projects, so I think it's a Server side setting that the Admins are okay with, yes some Projects make the changes immediately but not all do. |
Send message Joined: 5 Jul 11 Posts: 990 Credit: 376,143,149 RAC: 0 |
Just curious to see if anyone has really studied whether there's an appreciable gain in running >1 WU/GPU? It depends on your CPU speed related to your GPU speed. I have a machine with a fast CPU and a slow GPU, it gets no benefit from two WUs per GPU. I have a machine with a slow CPU and a fast GPU, it certainly benefits from two WUs per GPU. Presumably because with only one WU per GPU, when it gets to a part of the WU that requires come calculations over on the CPU, the GPU is sat idle waiting for it. With a good CPU, that time is minimal. With a slower CPU, having a second WU the GPU can do while it's waiting means it's in use more often. All I did was try it with 1 WU per GPU, note how long it took to complete a WU, then try two WUs, and note how long those WUs took (obviously taking exactly twice as long would mean no change, as it's doing two at once). Then I even tried 3 and 4 WUs. In my setups, I usually find 2 WUs is an improvement, and 3 4 etc makes no change. But I have old computers with good GPUs - as I spend all the cash on GPUs. |
Send message Joined: 7 May 14 Posts: 57 Credit: 206,540,646 RAC: 2 |
hi all made vid on youtube for multiple instances instruction's and at full load on a Radeon VII RADEON VII GIGABYTE// 3 Instances_ Milkyway@home WUs BOINC_ 3_instances https://www.youtube.com/watch?v=4xKy9wGKmz4 all the best and welcome to earth |
©2024 Astroinformatics Group