Message boards :
Number crunching :
Did the work unit limit change?
Message board moderation
Previous · 1 · 2 · 3 · 4
Author | Message |
---|---|
Grant (SSSF) Send message Joined: 19 Aug 99 Posts: 13736 Credit: 208,696,464 RAC: 304 |
A system with NVIDIA & iGPU the 100 tasks may only be assigned to NVIDIA leaving the iGPU idle. Now that system can download 200 GPU tasks, but they may all be assigned to NVIDIA leaving the iGPU still idle. Yep. All the posts I've seen so far have pointed out how each GPU now has 100WU as opposed to sharing only 100WU in total. Other than some inactive GPUs being used when working out the limit (which means the system has more than 100WU per active GPU) i hadn't seen any instances where inactive or much slower ones were getting work, and faster or active GPUs weren't getting their 100WUs. EDIT- Thinking about it I think the ideal way to do what they are trying to implement would be to take the project limit and split it across the different vendor GPU's. Probably based on their processing rate. So a System with a faster NVIDIA card and an iGPU might have a limit of like 70 for the NVIDIA & 30 for the iGPU. If they were to do that then they will have to raise the GPU limits significantly. The problem with the limits is the GPUs running out of work even with very short outages. Splitting the allocation between GPUs will make it even worse. I'm a fan of the KISS (Keep It Simple Stupid) principle- splitting the allowance across multiple cards & multiple vendors makes for considerable complexity, as opposed to just allowing the limit to be changed from x GPU WUs per system to x WUs per GPU per system. No need to worry about vendors or percentages. Just sort out the issue that is resulting in inactive GPUs being used in the allocation of work. Grant Darwin NT |
HAL9000 Send message Joined: 11 Sep 99 Posts: 6534 Credit: 196,805,888 RAC: 57 |
A system with NVIDIA & iGPU the 100 tasks may only be assigned to NVIDIA leaving the iGPU idle. Now that system can download 200 GPU tasks, but they may all be assigned to NVIDIA leaving the iGPU still idle. My 2nd example with the slower iGPU getting all of the work may never happen currently. I just thought it was a good example of demonstrating how things could go wrong. EDIT- Fixing the db and scripts so it doesn't get grumpy & crash would probably be easier than implementing a split limit system. Once they fix the problem of applying the limits to the type of GPU's the issue with allocating tasks for inactive GPU's will go away. If they can't make it work as expected hopefully they will remove this code so the BOINC servers actually follow the limits set by the project admins. SETI@home classic workunits: 93,865 CPU time: 863,447 hours Join the [url=http://tinyurl.com/8y46zvu]BP6/VP6 User Group[ |
Grant (SSSF) Send message Joined: 19 Aug 99 Posts: 13736 Credit: 208,696,464 RAC: 304 |
Fixing the db and scripts so it doesn't get grumpy & crash would probably be easier than implementing a split limit system. Not having the limits at all would be the best option. Having them per GPU is the next best. I suspect the problems with the database are hardware more than software related, and if the requirements were made known then I'd expect a quick fundraiser would get what's needed. But as there hasn't been any mention of what the cause or fix is, not much we can do to help. Grant Darwin NT |
Josef W. Segur Send message Joined: 30 Oct 99 Posts: 4504 Credit: 1,414,761 RAC: 0 |
The assumption that the BOINC changeset was the only cause for the limits changing is probably incorrect. As Claggy noted in his msg 1515079 near the beginning of this thread, he reported the issue on the boinc_dev mailing list. Eric Korpela's reply was: I changed config_aux.xml limit to have the <per_proc/> option set. Checking the haveland graphs indicates the change has had essentially no effect on turnaround time nor how many tasks are "in progress", so there's no perceptible downside as far as project load goes. I think the overall effect is good. Joe |
HAL9000 Send message Joined: 11 Sep 99 Posts: 6534 Credit: 196,805,888 RAC: 57 |
The assumption that the BOINC changeset was the only cause for the limits changing is probably incorrect. As Claggy noted in his msg 1515079 near the beginning of this thread, he reported the issue on the boinc_dev mailing list. Eric Korpela's reply was: I have been watching that since the change. I suspect doing AP on GPU's has helped a bit with keeping the number of MB tasks out in the field down. It also looks like more data is being moved over to be split. so there might be more AP for the weekend to keep the GPU's from going to town on MB's. SETI@home classic workunits: 93,865 CPU time: 863,447 hours Join the [url=http://tinyurl.com/8y46zvu]BP6/VP6 User Group[ |
Grant (SSSF) Send message Joined: 19 Aug 99 Posts: 13736 Credit: 208,696,464 RAC: 304 |
Checking the haveland graphs indicates the change has had essentially no effect on turnaround time nor how many tasks are "in progress", so there's no perceptible downside as far as project load goes. I think the overall effect is good.Joe In my case the turnaround time for GPU WUs has gone from around 8 hours to 16 hours. While a significant increase (it's doubled), in absolute terms it's sweet bugger all (the turn around time for my CPUs are 43hrs & 216hrs). Grant Darwin NT |
petri33 Send message Joined: 6 Jun 02 Posts: 1668 Credit: 623,086,772 RAC: 156 |
I could. I will not. I'll take the mixture. (going childish) Would You be asking if I had 240k. -- Now I'm a bit down. To overcome Heisenbergs: "You can't always get what you want / but if you try sometimes you just might find / you get what you need." -- Rolling Stones |
petri33 Send message Joined: 6 Jun 02 Posts: 1668 Credit: 623,086,772 RAC: 156 |
I could.[/quote] Thanks Tim! (looked just at your tasklist -- 800 MB ;) ) Br Petri To overcome Heisenbergs: "You can't always get what you want / but if you try sometimes you just might find / you get what you need." -- Rolling Stones |
Grant (SSSF) Send message Joined: 19 Aug 99 Posts: 13736 Credit: 208,696,464 RAC: 304 |
Thinking of the increased database load that larger caches would cause, I notice that I've still got 13 Enhanced WUs still sitting in my account. I'd expect clearing all those older WUs out would have to help the database considerably. Grant Darwin NT |
ExchangeMan Send message Joined: 9 Jan 00 Posts: 115 Credit: 157,719,104 RAC: 0 |
Thinking of the increased database load that larger caches would cause, I notice that I've still got 13 Enhanced WUs still sitting in my account. I'd expect clearing all those older WUs out would have to help the database considerably. I still have 34 of them, all from last year in May and June. |
©2024 University of California
SETI@home and Astropulse are funded by grants from the National Science Foundation, NASA, and donations from SETI@home volunteers. AstroPulse is funded in part by the NSF through grant AST-0307956.