Message boards :
Number crunching :
Server Updates and Status
Message board moderation
Author | Message |
---|---|
Send message Joined: 30 Aug 07 Posts: 2046 Credit: 26,480 RAC: 0 |
This is for discussion of the latest news item. We're going to be splitting MilkyWay@Home into two separate applications/backends, one for GPUs and one for regular processors. We've been having some complaints from RPI about our bandwidth usage (it's up to around 250gb/week), so this is in part to address that, but also to make sure work stays smoothly flowing and the server doesn't get bogged down. We're going to keep the regular milkyway application doing what it has been -- global optimization methods such as particle swarm, genetic search, etc. The GPU version of milkyway is going to farm more complex calculations to the application (gradient/hessian/line search calculations) which will let them do more "local" searches. This should increase the WU size of the GPU application by 300-1000 times what they currently are, significantly decrease server load, and also reduce the network usage of these clients. |
Send message Joined: 12 Nov 07 Posts: 2425 Credit: 524,164 RAC: 0 |
Sounds good. This is still going to share the current server but run as 2 different functions then? Doesn't expecting the unexpected make the unexpected the expected? If it makes sense, DON'T do it. |
Send message Joined: 30 Aug 07 Posts: 2046 Credit: 26,480 RAC: 0 |
Sounds good. This is still going to share the current server but run as 2 different functions then? Yeah I think that's how we're going to do it. The server should be able to handle it if the GPU request are significantly cut down. |
Send message Joined: 11 Dec 07 Posts: 11 Credit: 95,037,645 RAC: 0 |
good news ! |
Send message Joined: 30 Aug 07 Posts: 2046 Credit: 26,480 RAC: 0 |
One thing we'd like to do, in a sense have it as a separate project (but with the same home page and forums), that way we could have top lists for milkyway regular and milkyway GPU, without them stepping on each others toes. I'm not sure if we can do this having GPU milkyway as just a separate application or not. |
Send message Joined: 18 Nov 07 Posts: 280 Credit: 2,442,757 RAC: 0 |
The GPU version of milkyway is going to farm more complex calculations to the application (gradient/hessian/line search calculations) which will let them do more "local" searches. This should increase the WU size of the GPU application by 300-1000 times what they currently are, significantly decrease server load, and also reduce the network usage of these clients. While it's good to hear that you'll be offloading the servers, could you comment on the science impact of these changes? |
Send message Joined: 10 Aug 08 Posts: 218 Credit: 41,846,854 RAC: 0 |
I'm not sure if we can do this having GPU milkyway as just a separate application or not. Your plan sounds good and maybe the answer. One thing though, will you have any controls in place to avoid those using GPU's to only do those particular workunits? With the way things are now its a free-for-all trying to get work. I guess what I am trying to say is that will your backend be able to tell that the client is running the GPU application and only send work appropriate as well as knowing that the CPU applications are requesting their work specific loads? PS.. While you were away there were some suggestions that some of the participants would like to donate necessary equpiment, etc to help the project out. Maybe you can look into this and see whether or not it is feasible. Glad to see you back and feeling better after your bout with the flu and everything. |
Send message Joined: 9 Nov 07 Posts: 151 Credit: 8,391,608 RAC: 0 |
One thing we'd like to do, in a sense have it as a separate project (but with the same home page and forums), that way we could have top lists for milkyway regular and milkyway GPU, without them stepping on each others toes. I think this approach would satisfy a lot of users! Lets hope it can be done. |
Send message Joined: 30 Aug 07 Posts: 2046 Credit: 26,480 RAC: 0 |
I guess what I am trying to say is that will your backend be able to tell that the client is running the GPU application and only send work appropriate as well as knowing that the CPU applications are requesting their work specific loads? We're probably going to have a separate project so only GPU clients can sign up for the GPU version, and we can keep credits separate. PS.. While you were away there were some suggestions that some of the participants would like to donate necessary equpiment, etc to help the project out. Maybe you can look into this and see whether or not it is feasible. The project PIs are looking into this, funding things are really not in my job description :) |
Send message Joined: 22 Nov 07 Posts: 285 Credit: 1,076,786,368 RAC: 0 |
|
Send message Joined: 22 Nov 08 Posts: 136 Credit: 319,414,799 RAC: 0 |
Will we be able to do GPU and CPU work on the same system? 4870 GPU 4870 GPU |
Send message Joined: 22 Nov 08 Posts: 136 Credit: 319,414,799 RAC: 0 |
This should increase the WU size of the GPU application by 300-1000 times what they currently are, significantly decrease server load, and also reduce the network usage of these clients. With the granted credit be 300-1000 times more per work unit. ;) I knew it was going to be asked. I wanted to do it first. <EDIT> I guess I should have read the news first. You've already answered this.. 4870 GPU 4870 GPU |
Send message Joined: 30 Aug 07 Posts: 2046 Credit: 26,480 RAC: 0 |
Will we be able to do GPU and CPU work on the same system? Should be able to, yes. That's our goal at any rate. |
Send message Joined: 30 Aug 07 Posts: 2046 Credit: 26,480 RAC: 0 |
Something like that, yes. |
Send message Joined: 16 Feb 09 Posts: 109 Credit: 11,089,510 RAC: 0 |
That’s an exciting idea. It could go a long way towards keeping different types of volunteers supplied with work. I’m sure that the extra science to be done on the new GPU app will be a welcome benefit too. Any time frame on implementing this? |
Send message Joined: 1 Dec 08 Posts: 139 Credit: 8,721,208 RAC: 0 |
Travis wrote: The GPU version of milkyway is going to farm more complex calculations to the application (gradient/hessian/line search calculations) which will let them do more "local" searches. This should increase the WU size of the GPU application by 300-1000 times what they currently are, significantly decrease server load, and also reduce the network usage of these clients. Sounds like a fine (and very clever) idea! |
Send message Joined: 12 Apr 08 Posts: 621 Credit: 161,934,067 RAC: 0 |
We're probably going to have a separate project so only GPU clients can sign up for the GPU version, and we can keep credits separate. It would almost certainly be easier to have two separate projects with two URLs. Otherwise I do not see how you would be able to keep all the "stuff" separated. I mean, you could then have the Milky Way project and the Andromeda project ... :) Or maybe the Large Magellanic Cloud project? |
Send message Joined: 4 Oct 08 Posts: 1734 Credit: 64,228,409 RAC: 0 |
An excellent idea Travis, and it would relieve the server load. The new GPU client and WU would, I assume, be looking for much more fine detail in the science than the MW broad based structures the CPU client looks for? Go away, I was asleep |
Send message Joined: 2 Feb 09 Posts: 1 Credit: 1,143,453 RAC: 0 |
Will we be able to do GPU and CPU work on the same system? My machine has a two Nvidia and one AMD graphics cards. Currently I crunch MW on the HD3870 and the CPU, and crunch GPUGRID on the GTX260s. Will it become possible for me to crunch MW on both types of GPU at the same time, and the CPU as well? |
Send message Joined: 27 Feb 09 Posts: 45 Credit: 305,963 RAC: 0 |
Nice to see I can use my measly nVidia card for this project especially in OSX. Mars rules this confectionery war! |
©2024 Astroinformatics Group