Message boards :
Number crunching :
Panic Mode On (77) Server Problems?
Message board moderation
Previous · 1 . . . 12 · 13 · 14 · 15 · 16 · 17 · 18 . . . 22 · Next
Author | Message |
---|---|
Grant (SSSF) Send message Joined: 19 Aug 99 Posts: 13727 Credit: 208,696,464 RAC: 304 |
You must have some kind of out of this world GPUs because that is common knowledge the NVIDIA GPUs are incredible slow when process VLAR´s. You could verify that in several threads in this forum. That was the pre-Fermi GPUs. My GTX 560Ti ran 2 at a time no problems when they were coming through. Most of them didn't run much longer than the usual long running WUs, but a few of them did blow out to over an hour, which was still more than 3 times faster than my E6600 can process them. And even my GTX460 didn't get bogged down with them either- but i am running the Lunatics optimised applications. Running the stock application it could be a whole different kettle of fish. Grant Darwin NT |
bill Send message Joined: 16 Jun 99 Posts: 861 Credit: 29,352,955 RAC: 0 |
I have previously posted my times for work units. My Nvidia gpus process vlars faster than my cpus. Just like astopulse wus process faster on qualified gpus than cpus. I can't help it if you and others are working off insufficient data. I suggest you and others accept the facts that software improves, hardware improves, and old assumtions based on out of date data no longer apply to newer hardware. and no, Wiggo, I've never had the problem that you refer to. |
juan BFP Send message Joined: 16 Mar 07 Posts: 9786 Credit: 572,710,851 RAC: 3,799 |
I have previously posted my times for work units. I realy dont agree with you, i have a lot of NVIDIA GPUs from the 560 to the 690 and use the lates x41z builds, in ALL GPUS the VLAR process very slow in comparation of a normal WU, as I know there are no software or hardware more updated. I think you make a confusion, we are talking about the time to process a WU in a GPU not in a CPU, of course if you process a VLAR on a GPU is faster than process on a CPU, but that is not the point. On a 580/670 a normal WU takes 10 min to process, a VLAR takes + 1.2 hours. That is the reason why the VLARS are not sended to NVIDIA. Probaily the same VLAR will take 1.5 hr to process on a normal CPU (not top of the line) So is clear the waste of time/resources spended if you try to process the VLARs on a NVIDIA. |
bill Send message Joined: 16 Jun 99 Posts: 861 Credit: 29,352,955 RAC: 0 |
I have previously posted my times for work units. Your opinion. I don't agree. On my machine it takes less resources and time to run a vlar on my gpu than on my cpu. |
juan BFP Send message Joined: 16 Mar 07 Posts: 9786 Credit: 572,710,851 RAC: 3,799 |
In this case you need to show us who to do that, a lot of people will be interested in your solution. As i know they are a lot who try to find this for years. Whow to run a VLAR fast on a Nvidia GPU. |
bill Send message Joined: 16 Jun 99 Posts: 861 Credit: 29,352,955 RAC: 0 |
Define fast. I can tell you that a GTX450, GTX460, GTX560Ti is considerably faster using Win 7 64 with all patches installed, Lunatics Win64 v0.40, Boinc 6.10.58, 6.10.60, 7.0.25, and 7.0.31 and my app_info set to run two instances per gpu than running the same vlar wu on either my e6600 or my q9550 cpu's. Anything else would be an unsubstantiated guess on my part. But the above does give lie to the vlars don't run on Nvidia gpu's. Others have seen similar results on their machines, just check past posts. |
Horacio Send message Joined: 14 Jan 00 Posts: 536 Credit: 75,967,266 RAC: 0 |
For SETI I have a host using a GTX680, one using a GT430 and another using a GT9500, all of them with Lunatics apps... Recently, when the vlars were sent to nvidia GPUs, I had to abort all of them, not due to speed, neither because they were failling, but simply because while they were beeing crunched the hosts become so unresponsive and the video was lagging so baldly that I was not able to use them for anything else... The host with GTX680 was the less affected, but in this host I run 2 VMs and the process running on those VMs were failling with I/O timeouts and other weird errors... That's the main reasson for the exclussion of vlars on nividia GPUs, if it were just a thing of performance for sure they wouldnt had spent even a minute to set an exclussion mechanism... Anyway, I agree that adding an option to opt in/out for VLARS on GPUs, may be a better path for us, but more complex to set and maintain on the server/lab side where they have more in hands that they can handle... |
Grant (SSSF) Send message Joined: 19 Aug 99 Posts: 13727 Credit: 208,696,464 RAC: 304 |
But the above does give lie to the vlars don't run on Nvidia gpu's. I'm not aware of anyone saying they won't run; but when people first started crunching with GPUs & the stock applications, a VLAR on the GPU would make the machine in question extremely sluggish, sometimes to the point of being unresponsive. For my GTX650Ti & GTX460 running the optimised applications, that isn't the case. But if running stock it may well still be the case. With v7 making use of optimised code for the stock applications, VLARs on a NVidia GPU or not will most likely just become another bit of history. Grant Darwin NT |
Slavac Send message Joined: 27 Apr 11 Posts: 1932 Credit: 17,952,639 RAC: 0 |
|
Slavac Send message Joined: 27 Apr 11 Posts: 1932 Credit: 17,952,639 RAC: 0 |
|
bill Send message Joined: 16 Jun 99 Posts: 861 Credit: 29,352,955 RAC: 0 |
Wasn't the 9500 the old nvidia architecture that really did puke on vlars? The GTX430 probably doesn't have enough power to do well, my gtx450 lags just a little bit if I have too many tabs open in my web browser, but that machine just crunches now, if I even bother to turn it on. Heat, power, summer, money, wife. Had to turn some stuff off till winter. The 680 you would think wouldn't be a problem, but I read somewhere around here that the gtx6xx architecture had been crippled in some way and that the equivalent gtx5xx card actually out performed it in some areas. If I got hold of a gtx680 I'd like to see how it does work. Maybe for Christmas. Otherwise don't know what to tell you. |
bill Send message Joined: 16 Jun 99 Posts: 861 Credit: 29,352,955 RAC: 0 |
We can only hope. |
Horacio Send message Joined: 14 Jan 00 Posts: 536 Credit: 75,967,266 RAC: 0 |
Wasn't the 9500 the old nvidia architecture that really did puke on Its a pre-fermi, so obviously it was completely expected to puke on the vlars, but there are a lot of prefermi GPUs like mine out there... The GTX430 probably doesn't have enough power to do well, my gtx450 Well, its a low profile, low cost, low power GPU, but its fermi and while is the little sister of the 400 series, it give me the same RAC as the 8 HT cores of the i7-2600 in which it runs (or more)... The 680 you would think wouldn't be a problem, but I read somewhere around here that the gtx6xx architecture had been crippled in some way and that the equivalent gtx5xx card actually out performed it in some areas. If I got hold of a gtx680 I'd like to see how it does work. Maybe for Christmas. Otherwise don't know what to tell you. The only thing crippled on the high end 600 series is the double precission performance... (the mid and low end 600 GPUs are the ones that have some other things crippled, mostly memory bus width) Anyway, I think that in this host, the vlars were just the drop that flooded the glass... I guess that if the host were just a cruncher I might not been noticed anything seriuos... The point here is that filtering vlars per GPU model is almost impossible withoud doing serious and complex recoding of the schedullers and worst with a lot of access to databases to see what GPU the requester has (the client software asks work for the NVIDIA, ATI or CPU, so the scheduller will need to query the database on each RPC to know what GPU the client has..) The current workaround was intended as a temporal fix, with the hope to get it solved at the apps level... but, the lack of time and resources along with other priorities, made the "temporal" thing a bit longer than expected... |
S@NL Etienne Dokkum Send message Joined: 11 Jun 99 Posts: 212 Credit: 43,822,095 RAC: 0 |
Is anyone still having issues with Auto Errored tasks? Basically tasks that error out before they're even downloaded or process, usually with impossible deadlines? yes, every now and then I get 10 or so WU's with a deadline of 5 minutes... Not often the last 3-4 days anymore |
S@NL Etienne Dokkum Send message Joined: 11 Jun 99 Posts: 212 Credit: 43,822,095 RAC: 0 |
started accepting new work again this morning. Got 198 GPU tasks in half an hour, so DL is working pretty good. Shame they're still all shorties ... |
Grant (SSSF) Send message Joined: 19 Aug 99 Posts: 13727 Credit: 208,696,464 RAC: 304 |
Is anyone still having issues with Auto Errored tasks? Basically tasks that error out before they're even downloaded or process, usually with impossible deadlines? Just had a look, and nothing on my task list since the 3rd. Sent 3 Oct 2012 | 22:54:50 UTC, Time reported/Deadline 3 Oct 2012 | 23:01:09 UTC, Status Timed out - no response. But i haven't had many resends since then either. At the peak of the problems with uploading/getting work, probably 1 in 5 to 1 in 7 requests that resulted in work would have been resends. Grant Darwin NT |
bill Send message Joined: 16 Jun 99 Posts: 861 Credit: 29,352,955 RAC: 0 |
That's okay, I don't have a problem with the lab boys "fix". I just have a problem with broad brushes that paint an incorrect picture. |
Bernie Vine Send message Joined: 26 May 99 Posts: 9954 Credit: 103,452,613 RAC: 328 |
Is anyone still having issues with Auto Errored tasks? Basically tasks that error out before they're even downloaded or process, usually with impossible deadlines? This happens where VLARS that were sent to your CPU failed to be delivered for some reason, they were then resent when your GPU requested tasks. As there is some speculation as to whether VLARS work well with NVIDIA GPUS they are not supposed to be sent so they time out after a few minutes. I know a lot of people are unhappy about this and hopefully something will be done. Perhaps. |
Grant (SSSF) Send message Joined: 19 Aug 99 Posts: 13727 Credit: 208,696,464 RAC: 304 |
As there is some speculation as to whether VLARS work well with NVIDIA GPUS they are not supposed to be sent so they time out after a few minutes. From what i've seen on the few of mine i've looked at it's always after 5 min. So i figure it's the next Scheduler request where the Scheduler finds out they've gone to the wrong place, so it just times them out. Grant Darwin NT |
robertmiles Send message Joined: 16 Jan 12 Posts: 213 Credit: 4,117,756 RAC: 6 |
Wasn't the 9500 the old nvidia architecture that really did puke on I've read elsewhere that the main crippling of the GTX680 seems to be the much smaller fraction of its GPU cores that can do double precision calculations. |
©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.