Message boards :
Number crunching :
guppie on NVIDIA cards
Message board moderation
Previous · 1 · 2
Author | Message |
---|---|
Mike Send message Joined: 17 Feb 01 Posts: 34258 Credit: 79,922,639 RAC: 80 |
the read me I have shows ... I know. I wrote this for 3330 back then. Things have slightly changed since. I wanted to update it but had not much time to do it. With each crime and every kindness we birth our future. |
Stubbles Send message Joined: 29 Nov 99 Posts: 358 Credit: 5,909,255 RAC: 0 |
With all the talk about rescheduling WUs and not having guppie's run on NVIDIA cards I would just like to share my experience. Hey Bill, I had a look again at your initial post and I don't think I answered directly your 1st part ...even with my looooong replies! I think Jeff did a good job at explaining with his #s...so here's my "short" answer (for me lol) with your #s mostly! In your comparison of guppi and non-guppi, you are missing the time it takes for the CPU to process non-guppi. ( I will continue to use your terminology if we can leave Arecibo.VLARs out of the picture, otherwise it makes things too fuzzy ). On the GTX 750 Ti, you already know the times for: non-guppi (20mins) < guppi (30 mins), and the ratio is 3:2 in favor of non-guppiIf your AMD 8cores are like my Xeon 8cores (with HT enabled) and you're using a Lunatics CPU app, you'll get: non-guppi (?) > guppi (3hrs) So to keep this short answer fairly short, do the math for set of calculations for 2 PCs over 2 days, where: The server sends you: - on day 1: 2 non-guppi for every guppy (2:1) - on day 2: 1 non-guppi for every guppi (1:1) and: - PC#1: non-guppi preferred on GPU - PC#2: guppi preferred on GPU If your AMD cores are like my Xeon cores, "non-guppi preferred" on GPU will always win! |
Stubbles Send message Joined: 29 Nov 99 Posts: 358 Credit: 5,909,255 RAC: 0 |
In order to keep my last post short, I omitted my scenario with ideal ratios. What's very different with my #s (as compared to Bill's) is that I get on the CPU cores: non-guppi (~3hrs) > guppi (2+hrs), and the ratio is almost 3:2 in favor of guppi On my rigs, if I had to process 60 guppis and 60 non-guppi, it would be faster if I processed only guppi on the CPU and whatever else on the GPUs. Fortunately, with the ratios being 3:2 for both the CPU and GPU, it makes the results very convenient if we also assume 6cores in use and exactly 2hr/guppi. In 20 hrs, the 6cores would crunch 60 guppi, and My sole GTX 750 Ti would process 60 non-guppi. Do it the other way around and you need 30 hrs to process the 60 non-guppi on 6cores (assuming 3hr/non-guppi), and also 30 hrs to process 60 guppi on GPU. Any other mix will fall between 20hrs and 30hrs. Obviously, this isn't the perfect scenario since: - we never know what ratio of guppi to non-guppi the server will send; and - there's those slow everywhere Arecibo-VLARs I omitted to keep things simple! Any more Qs/concerns about the advantage of transferring: non-guppi CPU-assigned tasks to the GPU(s) for any number of GTX 750 Ti? ...as a general rule no matter what the server sends (...as long as the ratio is anywhere between 10:1 and 1:10)! [edit]Obviously a cap of a few hundred tasks assigned to each GPU is highly recommended![/e] |
Bill G Send message Joined: 1 Jun 01 Posts: 1282 Credit: 187,688,550 RAC: 182 |
I would love to do just that. Perhaps stock was not exactly the right term, I just meant that I added nothing to the download. SETI@home classic workunits 4,019 SETI@home classic CPU time 34,348 hours |
qbit Send message Joined: 19 Sep 04 Posts: 630 Credit: 6,868,528 RAC: 0 |
After a few month break I'm back crunching since a few days and I have problems with those Guppies also. I am running 2 tasks at a time on my 750 using the r3472 Open CL build. With the suggested command line from readme I can do a guppi in ~1 hour but the system seems to get a lot of stress, extremly unresponsible with lags up to a minute. I don't like that. When adding use_sleep everything runs fine, no lags, lower temps, BUT guppies take double the time, ~2 hours. Yesterday I changed the commandline to the one which Mike recommended in this thread but that doesn't seem to help either. Guppies seem to take even longer now, some more then 3 hours! Here's an example: http://setiathome.berkeley.edu/workunit.php?wuid=2209251262 Any suggestions? |
Stubbles Send message Joined: 29 Nov 99 Posts: 358 Credit: 5,909,255 RAC: 0 |
Hey qbit, After a few month break I'm back crunching since a few days and I have problems with those Guppies also...Here's a list of options in order of priority (for me): Option 1: as a volunteer tester, help Raistmer test his latest NV_SoG release on main project ( since guppis are likely to be the future standard tasks ) Option 2: connect to S@H beta and crunch the new 4-bit file format. Option 3: do option 1 & 2 in whatever order Raistmer suggests. Option 4: transfer guppis to CPU, and non-vlars to GPU (PM mee for my semi-automated batch file, or wait for Mr Kevvey's: (see thread ...swap...) and help test scripts Option 5: revert to gold standard for GTX 750 Ti: cuda50 with 2or3WU on GPU Option 6: do option 5 and then option 4 (for max throuput) Option 7: keep playing with NV_SoG commanline Option 8: suck it up and hope that Petri's future apps will be the answer to guppis Any Qs? RobG ;-) |
Mike Send message Joined: 17 Feb 01 Posts: 34258 Credit: 79,922,639 RAC: 80 |
After a few month break I'm back crunching since a few days and I have problems with those Guppies also. Did you try -spike_fft_thresh 2048 already ? You can also add -period_iterations_num 20 or 10 but this might increase screen lags again. Don`t forget the read me were written whilst guppie tasks wasn`t released. So fine tuning is necessry especially on mid range GPU`s. With each crime and every kindness we birth our future. |
qbit Send message Joined: 19 Sep 04 Posts: 630 Credit: 6,868,528 RAC: 0 |
Hey qbit, Hi Rob! Options 1-3: Sorry, but no, for reasons which I won't explain here. Option 4: I don't do CPU crunching on this machine. It has an E5300 dual core and I guess I need that to feed the GPU. With use_sleep I might be able to crunch with one core but to be honest I don't think that's worth it. Option 5: maybe Option 7: will do that Option 8: Yeah, I saw some of Petris results and IMO this looks very promising. But I guess it will take some time before those apps are released to the public.
Hey Mike! Not yet, but I will try! I don't mind a little lag, since it's a dedicated cruncher, I just don't like those really heavy lags. Shall I change -sbs 256 to -sbs 384 also? |
Raistmer Send message Joined: 16 Jun 01 Posts: 6325 Credit: 106,370,077 RAC: 121 |
After a few month break I'm back crunching since a few days and I have problems with those Guppies also. -use_sleep + 2-3 simultaneous tasks (additionally to Mikes' proposals). Or go anonymous platform and install r3486. SETI apps news We're not gonna fight them. We're gonna transcend them. |
qbit Send message Joined: 19 Sep 04 Posts: 630 Credit: 6,868,528 RAC: 0 |
Since yesterday I run this commandline: -use_sleep -sbs 384 -spike_fft_thresh 2048 -period_iterations_num 20 -tune 1 64 1 4 -oclfft_tune_gr 256 -oclfft_tune_lr 16 -oclfft_tune_wg 256 -oclfft_tune_ls 512 -oclfft_tune_bn 32 -oclfft_tune_cw 32 Seems to do the trick, most guppies run around 6k seconds now, which looks like a nice speed up!
Not sure about running 3 tasks at one, since my card is just a non-Ti GTX 750 with 1 Gig RAM. But I may check out r3486 soon! |
qbit Send message Joined: 19 Sep 04 Posts: 630 Credit: 6,868,528 RAC: 0 |
So today I switched from r3472 to r3486. The latter one utilizes the CPU much more but also seems to speed up Guppies by quite a bit. Check this: ar 0.007531, crunched with r3486, time 1h 07m 01s, CPU time 0h 36m 22s https://setiathome.berkeley.edu/result.php?resultid=5046235465 ar 0.007423, crunched with r3472, time 1h 41m 10s, CPU time 0h 1m 03s https://setiathome.berkeley.edu/result.php?resultid=5045244498 Of course this needs more testing, just running r3486 for a few hours yet, but so far it looks promising. |
Raistmer Send message Joined: 16 Jun 01 Posts: 6325 Credit: 106,370,077 RAC: 121 |
So today I switched from r3472 to r3486. The latter one utilizes the CPU much more but also seems to speed up Guppies by quite a bit. Thanks for testing. I hope next rev will have reduced CPU usage with sleep enabled again. SETI apps news We're not gonna fight them. We're gonna transcend them. |
Stubbles Send message Joined: 29 Nov 99 Posts: 358 Credit: 5,909,255 RAC: 0 |
Let me know if you are interested in doing different setups for your 3 PCs, or if you have any Qs or concerns about doing so.I would love to do just that. Hey Bill, Have you found the time to convert one or two of your rigs with 2x 750 Ti? Rob :-) |
Bill G Send message Joined: 1 Jun 01 Posts: 1282 Credit: 187,688,550 RAC: 182 |
Let me know if you are interested in doing different setups for your 3 PCs, or if you have any Qs or concerns about doing so.I would love to do just that. No I have not had any time....in fact I will be shutting them all down in the morning as I am going on vacation and I do not want to leave the AC on in the computer room the entire time I am gone. With the daily temps hitting over 32C I can not leave them alone. Let you know when I get back and perhaps have some time. (new HTT to set up and pack) SETI@home classic workunits 4,019 SETI@home classic CPU time 34,348 hours |
©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.