Message boards :
Number crunching :
Windows 10 - Yea or Nay?
Message board moderation
Previous · 1 . . . 129 · 130 · 131 · 132 · 133 · 134 · 135 . . . 163 · Next
Author | Message |
---|---|
tullio Send message Joined: 9 Apr 04 Posts: 8797 Credit: 2,930,782 RAC: 1 |
The BOINC Manager log sees both CUDA and OpenCl in the nVidia 382.53 driver. Now BOINC uses CUDA on SETI@home tasks, I shall launch Einstein@home GPU tasks as soon as SETI tasks are completed. Tullio |
Jimbocous Send message Joined: 1 Apr 13 Posts: 1855 Credit: 268,616,081 RAC: 1,349 |
EDIT- I just noticed you're still using the CUDA application on your video cards. Any particular reason? I was still hanging loose waiting for the release of the latest Lunatics installer, which I thought was forthcoming last fall. Just didn't feel like doing the manual edit on 5 machines, and was also hoping to see the MB v7 stuff removed so BOINC would quit complaining. Have a couple of other issues that a reinstall would resolve, as well. Will need to address this, as I'm recently experiencing a lot of trouble with Cuda 5.0 and the newer NVidia drivers +/or Win 10, for example these restarts, 159 SETI@home 06/03/17 12:17:29 AM task postponed 180.000000 sec: Cuda device initialisation failedwhich often can only be resolved by a reboot. Now that I have 3 of 5 crunchers shut down for the summer, it's probably time to start addressing such issues, especially as I know I'm sacrificing production capability. |
Jimbocous Send message Joined: 1 Apr 13 Posts: 1855 Credit: 268,616,081 RAC: 1,349 |
So i'd say you're yet to get the Creators update. Well, the pretty screens say otherwise, as did the downtime I watched while each of the 4 updated and I had to restore the NVidia drivers :) System Info is: Edition Windows 10 Pro Version 1703 OS Build 15063.413 But none of the (now 4) systems offer the option you described. I wonder if that's something that again is only offered if you have a metered connection selected, and is otherwise hidden? Wouldn't put it past them. |
Grant (SSSF) Send message Joined: 19 Aug 99 Posts: 13769 Credit: 208,696,464 RAC: 304 |
EDIT- I just noticed you're still using the CUDA application on your video cards. Any particular reason? The Beta 6 installer takes care of all of that. I went back & forth between CUDA & SoG on my C2D a few times, and didn't lose any work while doing so with the Beta 6 installer. Grant Darwin NT |
Grant (SSSF) Send message Joined: 19 Aug 99 Posts: 13769 Credit: 208,696,464 RAC: 304 |
So i'd say you're yet to get the Creators update. That is, odd. Grant Darwin NT |
Jimbocous Send message Joined: 1 Apr 13 Posts: 1855 Credit: 268,616,081 RAC: 1,349 |
EDIT- I just noticed you're still using the CUDA application on your video cards. Any particular reason? OK, thanks for that. Been meaning to ask. I'll track it down. |
Jimbocous Send message Joined: 1 Apr 13 Posts: 1855 Credit: 268,616,081 RAC: 1,349 |
So i'd say you're yet to get the Creators update. I wonder if that's something that again is only offered if you have a metered connection selected, and is otherwise hidden? Wouldn't put it past them. |
Grant (SSSF) Send message Joined: 19 Aug 99 Posts: 13769 Credit: 208,696,464 RAC: 304 |
I wonder if that's something that again is only offered if you have a metered connection selected, and is otherwise hidden? Wouldn't put it past them. Maybe. From memory, part of the Creators Update was meant to also allow wired connections to be set as metered. Grant Darwin NT |
Grant (SSSF) Send message Joined: 19 Aug 99 Posts: 13769 Credit: 208,696,464 RAC: 304 |
OK, thanks for that. Been meaning to ask. I'll track it down. Here it is, Mike's world downloads. Grant Darwin NT |
tullio Send message Joined: 9 Apr 04 Posts: 8797 Credit: 2,930,782 RAC: 1 |
All Einstein@home GPU tasks have errored out using OpenCL on the Creators Update Windows 10 and the 382.53 nVidia driver. Instead the SETI Beta tasks, now using OpenCL seem to work. Einstein@home GPU tasks on the Linux box work flawlessly using OpenCl. Tullio |
Jimbocous Send message Joined: 1 Apr 13 Posts: 1855 Credit: 268,616,081 RAC: 1,349 |
The Beta 6 installer takes care of all of that. Installed and running nicely since late last night on the first box, 8.20 Sog tasks so far all validate and no errors with 382.33 driver. Time to tune it a bit. I did recall from my Beta work what the heartburn was that I had with the SoG apps, but this probably isn't the place for that. Suffice it to say that since Cuda50 seem to no longer be a viable app, it's time to suck it up. |
Grant (SSSF) Send message Joined: 19 Aug 99 Posts: 13769 Credit: 208,696,464 RAC: 304 |
Installed and running nicely since late last night on the first box, 8.20 Sog tasks so far all validate and no errors with 382.33 driver. Time to tune it a bit. -tt 1500 -hp -period_iterations_num 1 -high_perf -sbs 1024 -spike_fft_thresh 4096 -tune 1 64 1 4 -oclfft_tune_gr 256 -oclfft_tune_lr 16 -oclfft_tune_wg 256 -oclfft_tune_ls 512 -oclfft_tune_bn 64 -oclfft_tune_cw 64 The 2 parameters that have the biggest impact- sbs 1024 will be OK for all of your video cards, those that don't have enough memory (some OpenCL weirdness about what % of total VRAM can actually be used) will drop that value down to one they can use. period_iterations_num 1 is OK for high end cards on higher end CPUs for a daily use machine. For a dedicated cruncher, even with lower end hardware (eg 750Ti) it's still good. If on your daily computer keyboard input & screen lag is too noticeable, change the value to 5 & see how it goes. If still annoying, try 10, then 15 etc if it's still an issue. If you get driver restarts, you'll probably need the TDR hack, just to increase the time before it decides there's an issue that it needs to recover from. Open regedit then - Navigate to - H_KEY_LOCAL_MACHINE \ System \ CurrentControlSet \ Control \ Graphics Drivers. Right click on the open field and click on New / DWORD (32bit) Value, Then name it: TdrDelay -> double click it -> Set the value to 8. Screen shots about half way down this page for TDR registry edit. EDIT- OpenCL on Nvidia needs 1 CPU core per GPU WU to do it's thing, and 1 WU per GPU is best. So 4 video cards, 4 cores required. The sleep function does reduce the CPU requirement, but it also knocks back output. I've personally found it worthwhile reserving 1 thread per GPU WU. The loss of CPU output is well and truly offset by the increase in GPU output. I've got HyperThreading on and don't even bother with reserving 1 thread from crunching for system use, it's usable with all cores working at 100%. Grant Darwin NT |
Jimbocous Send message Joined: 1 Apr 13 Posts: 1855 Credit: 268,616,081 RAC: 1,349 |
Actually have it set for that, except -tt 700 and no -hp, based on info from a tuning message a while back. This box is single 750ti on a C2Q, my temporary daily driver until I can get my 2nd Z600 back up. |
Grant (SSSF) Send message Joined: 19 Aug 99 Posts: 13769 Credit: 208,696,464 RAC: 304 |
Have you just set those values, as your current Valid and Pending results show the default values being used? Also your system is only showing 1 Processor, so only GPU crunching will be possible unless it shares that core with a CPU WU. Will really blow out the crunching times. Grant Darwin NT |
Jimbocous Send message Joined: 1 Apr 13 Posts: 1855 Credit: 268,616,081 RAC: 1,349 |
Yeah, just set that then had to back off due to restarts. System shows 1 CPU as three cores reserved for GPUs at this point. (cc_config <ncpus>1</ncpus>) |
Grant (SSSF) Send message Joined: 19 Aug 99 Posts: 13769 Credit: 208,696,464 RAC: 304 |
System shows 1 CPU as three cores reserved for GPUs at this point. (cc_config <ncpus>1</ncpus>) Ah, never seen it that way before. It makes it look like those threads aren't available to BOINC at all. I've always gone with app_config.xml Reserves a thread for each GPU WU, when not used by the GPU it's available for CPU crunching again. <app_config> <app> <name>setiathome_v8</name> <gpu_versions> <gpu_usage>1.00</gpu_usage> <cpu_usage>1.00</cpu_usage> </gpu_versions> </app> <app> <name>astropulse_v7</name> <gpu_versions> <gpu_usage>0.5</gpu_usage> <cpu_usage>1.0</cpu_usage> </gpu_versions> </app> </app_config> Grant Darwin NT |
Jimbocous Send message Joined: 1 Apr 13 Posts: 1855 Credit: 268,616,081 RAC: 1,349 |
System shows 1 CPU as three cores reserved for GPUs at this point. (cc_config <ncpus>1</ncpus>) Nice thing about doing it in cc_config is that it can be done on the fly, with changes being immediate and not requiring a shutdown/restart, unlike app_config changes. If you use BoincTasks, the change can be made there for each monitored machine. Just click a CPU task on the machine to be changed, then select Extra>Edit Config file. When you Apply, the client rereads on the fly and the change is immediate. Perhaps not as flexible as doing it in app-config, but still Nice. |
Brent Norman Send message Joined: 1 Dec 99 Posts: 2786 Credit: 685,657,289 RAC: 835 |
..... (cc_config <ncpus>1</ncpus>) I did that too on my AMD 4200 so that it will only run 1 CPU task - when it thinks the CPU is in high priority, it will take both cores and choke the system. It can't do that if it only has access to 1 CPU. Then just don't reserve a core, since one is free all ready. |
Grant (SSSF) Send message Joined: 19 Aug 99 Posts: 13769 Credit: 208,696,464 RAC: 304 |
Nice thing about doing it in cc_config is that it can be done on the fly, with changes being immediate and not requiring a shutdown/restart, unlike app_config changes. For app_config, all you need to do is Options, read config files. Your run times are still longer than i would have expected- from memory my C2D with the GTX 750Tis was doing WUs in around17min with SoG. Are you crunching any CPU WUs at the moment? What is the CPU utilisation in Task Manager? The way I read that ncpus setting is that it allows you to increase the number of WUs that can run on a single CPU core, not for reserving cores for other uses. Grant Darwin NT |
Jimbocous Send message Joined: 1 Apr 13 Posts: 1855 Credit: 268,616,081 RAC: 1,349 |
Nice thing about doing it in cc_config is that it can be done on the fly, with changes being immediate and not requiring a shutdown/restart, unlike app_config changes. Yeah, I'm taking a little different approach to sneaking up on this, as I did before when doing Beta. I started out Cuda50 running 3 tasks/GPU, as I have forever. Then ran 3/gpu Sog for a bit (hence the long run times). Now down to 2 tasks/gpu and have 1 AP running on the CPU I'm trying to get pushed through. The 2 GPU tasks running in parallel, 1 is with default tuning, the second is with your suggested tuning. Watching to see how that performs, then see how the box does with 2/ tuned, watch for restarts, and will later back that down to 1/ for comparison. When I did this last fall, I think I actually came out a bit ahead at 2/. The other thing I do is watch CPU and GPU core and memory controller usage, power and temps. SoG CPU% are in the low 90s (untuned job) and high 90s (tuned job). somewhat more mouse/kb lag than I'll want to tolerate. |
©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.