Message boards :
Number crunching :
new setup dual e5-2660
Message board moderation
Author | Message |
---|---|
Steve S Send message Joined: 19 Jan 13 Posts: 27 Credit: 6,274,024 RAC: 0 |
Currently it's running in single cpu mode until I get my 2nd cpu. It also has a GTX 660 ti. Anything I can do to increase performance other than making sure it's cool to keep turbo high? Some tasks come in at an estimated 4.5 hours while others come in at 3.5 hours. The GPU tasks are all around 5-10 minutes. Full specs: 2 x E5-2660 ASUS Z9PA-D8 ATX 8x Crucial Ballistix Sport 8GB DDR3 1600 ASUS GTX660 TI-DC2O-2GD5 GeForce GTX 660 Ti 2GB Corsair AX850 |
spitfire_mk_2 Send message Joined: 14 Apr 00 Posts: 563 Credit: 27,306,885 RAC: 0 |
Maybe run two gpu tasks on the 660ti, make sure you have two cpu cores free to feed the 660ti. The rest of cpu to do cpu tasks. |
Wiggo Send message Joined: 24 Jan 00 Posts: 34744 Credit: 261,360,520 RAC: 489 |
My GTX 660's (non Ti's) run 3 tasks each to be at their most productive, but YMMV. Cheers. |
Steve S Send message Joined: 19 Jan 13 Posts: 27 Credit: 6,274,024 RAC: 0 |
Cool, do I run another instance of boinc or what? Wiggo says I can run 3, is that dangerous (too much heat / crash)? The tasks say cuda 32, is that only 32 cuda cores because these have about 1300 cuda cores. |
spitfire_mk_2 Send message Joined: 14 Apr 00 Posts: 563 Credit: 27,306,885 RAC: 0 |
Cool, do I run another instance of boinc or what? Wiggo says I can run 3, is that dangerous (too much heat / crash)? The tasks say cuda 32, is that only 32 cuda cores because these have about 1300 cuda cores. cuda32 is CUDA version 3.2 Latest version of CUDA released by nVidia is 5.5 There are several ways to run multiple tasks on a gpu. I think the easiest is to install seti@home optimized application that Lunatics have made. Once you install it, there will be a file created called app_info.xml. While seti is not running, open this file in Notepad, find all instances of <count>1</count> using Find function in Notepad, change 1 to .5 if you want to run two tasks on the gpu, change 1 to .3 if you want to run three tasks on gpu. This is how I do it. Lunatics seti@home optimized application: http://www.arkayn.us/forum/index.php?action=tpmod;dl=cat3 The normal CUDA version for 6xx family is CUDA version 5 I think. My guess the reason you have cuda32 task is because when you first install seti@home, the seti system tests your system by running task of all cuda versions. When I first changed over to seti 7, I had cuda23, cuda32, cuda42, cuda5. Now that my system has been tested, I only get cuda42, cuda42 is the appropriate cuda version for my gtx460 video card. You have 660ti, I think the appropriate cuda version for you is cuda5, once testing is done, you should be receiving only cuda5 tasks. The testing takes a few days. |
juan BFP Send message Joined: 16 Mar 07 Posts: 9786 Credit: 572,710,851 RAC: 3,799 |
You could do that even with stock apps by the use of app_config.xml file if you use Boinc 7.0.64 or +, several examples avaiable in the forum, and that works in the similar way when you use the lunnatics installer. One tip, if you run more than 1 WU on your GPU don´t forget it will run a lot hotter than normal, so use a free to DL program like EVGA Precision, or MSI Afterburner (there are others similars from other brands but basicaly all do the same job) to keep you GPU fan at high speed, the standard settings are normaly not enoght for continous crunching. |
Steve S Send message Joined: 19 Jan 13 Posts: 27 Credit: 6,274,024 RAC: 0 |
So I'll have 32 threads and I'll try running 3 GPU tasks, how does this look? <app_config> <app> <name>setiathome_v7</name> <max_concurrent>32</max_concurrent> <gpu_versions> <gpu_usage>0.3</gpu_usage> <cpu_usage>0.01</cpu_usage> </gpu_versions> </app> </app_config> So there's no difference between cpu_usage 1 or .01, right? Will the GPU finish faster if the cpu_usage value is increased from .01 to 1? By the way, now that the gpu_usage is at .3, it starts 3 CUDA tasks and they finish between 14-26 minutes. But before running 1 CUDA task, it was finishing between 7 to 14 minutes. Is that normal? My GPU temperature is the same as 1 GPU task. |
juan BFP Send message Joined: 16 Mar 07 Posts: 9786 Credit: 572,710,851 RAC: 3,799 |
I sugest you to read this: http://setiathome.berkeley.edu/forum_thread.php?id=72626&postid=1407786 obviusly just clear the <max_concurrent> line because he don't do CPU crunching. It shows a compleate working file. You could use <max_concurrent>32</max_concurrent> to limit the number of WU to 32 but i see not need to do that here, since your cpu Usage will to that automaticaly for you and limit the number of concurrent task. I just not understand why you are trying to allow run up to 32 tasks if you host page say you have only 16 processors? The times you show are perfectly normal. My 670 crunch a WU normaly in 16 min running 2 MB at a time. I try 3 but at the end the video on the host start to get some lag and the times are almost the same, since now i run AP i alow only 1AP+1MB or 2 MB´s on it. I belive your GPU could run 3 MB at the time (or 1AP +2M), but you need to try and compare the times. Even if theory Intel's CPU´s don't need when crunching MB only, i leave a core free all the time to feed the GPU even when i run only MB. |
Steve S Send message Joined: 19 Jan 13 Posts: 27 Credit: 6,274,024 RAC: 0 |
because i'll have 32 threads eventually, in 24 hours So now with gpu_usage 0.3, the GPU tasks are taking 30 minutes each when before they were completing around 10 minutes, what's the point? |
juan BFP Send message Joined: 16 Mar 07 Posts: 9786 Credit: 572,710,851 RAC: 3,799 |
because i'll have 32 threads eventually, in 24 hours Sorry i still can´t understand that, max_concurrent is the maximum number of simultaneusly crunching WU on a determinate host, nothing related to a 24 hs period. Maybe i miss something due the translation. If you have a 16 processors on the CPU you must start only a maximum of 16 WU at a time (plus the ones on the GPU´s of course less the one needed to feed the GPU) unless you are trying to run 2 WU on each processor, that normaly slowdowns everything and most be avoided but YMMV, each host is unique. So now with gpu_usage 0.3, the GPU tasks are taking 30 minutes each when before they were completing around 10 minutes, what's the point? Yes. But to be sure, release 1 core to feed the GPU and re-run the test. You shold try 2 (0.5) and see the times. Take a look at the msg on the other post. |
Steve S Send message Joined: 19 Jan 13 Posts: 27 Credit: 6,274,024 RAC: 0 |
because i'll have 32 threads eventually, in 24 hours No. Your translation is bad. I will have 32 threads in the future. So if these GPU tasks need a CPU thread, then shouldn't I set cpu_usage to 1? |
juan BFP Send message Joined: 16 Mar 07 Posts: 9786 Credit: 572,710,851 RAC: 3,799 |
No. Your translation is bad. I will have 32 threads in the future. I belive then you have plans to add a new processor, even then you don´t need to worry about that, the boinc adjusts itself automaticaly to the new avaible processor count. So if these GPU tasks need a CPU thread, then shouldn't I set cpu_usage to 1? Yes and NO. No if they are MB (that uses very little processor support) Yes if they are AP. That´s why i say leave 1 free core for all the MB´s and one extra core for each AP. And the supplied app_config does that automaticaly for you by free 1 core as each AP starts. You just need to free the first core manualy, on your Boinc preferences page of the project... On multiprocessors, use at most ... set to leave a minimum 1 core free. But remember, there are some who say you don´t even need to free the first core when crunching only MB, as each host is unique and YMMV the best whay to be sure if your particular hosts need or not is test, run with and without the first free core avaiable on Intel CPUs (AMD sure needs) and look the results. You will find your optimal setting by yourself. You could even try by free 2 cores but normaly 1 is enought in Intel´s CPU´s, specialy the big and faster ones like your´s. |
Steve S Send message Joined: 19 Jan 13 Posts: 27 Credit: 6,274,024 RAC: 0 |
No. Your translation is bad. I will have 32 threads in the future. Thanks for the info. So I switched gpu_usage to .5 like you suggested last night, the results were giving me 20 minutes average, but that is still slower than when I had it on 1 GPU. I have switched gpu_usage back to 1 and seeing what the results are. My "multiprocessors, use at most" setting was leaving 1 core free. But now I set cpu_usage to 1 and gpu_usage to 1 and mulitporcessor setting to 100%. This will make me run 15 cpu tasks and 1 gpu task. |
juan BFP Send message Joined: 16 Mar 07 Posts: 9786 Credit: 572,710,851 RAC: 3,799 |
My "multiprocessors, use at most" setting was leaving 1 core free. But now I set cpu_usage to 1 and gpu_usage to 1 and mulitporcessor setting to 100%. This will make me run 15 cpu tasks and 1 gpu task. Thats fine, works as expected, but i still belive 2 WU or maybe 3 on the GPU is the optimal setting,the 660TI is a very fast GPU, 1 sure not. 1 core is about 6.25% of your CPU (1/16 processors) so a number or 93% or 94%, I not remember if the rounding is up or down, will free the first core. You could easely see what is the right one because when you free the core you will run 14 CPU task and 1 GPU task if is AP or 15 CPU + 1 GPU if is MB. |
Steve S Send message Joined: 19 Jan 13 Posts: 27 Credit: 6,274,024 RAC: 0 |
My "multiprocessors, use at most" setting was leaving 1 core free. But now I set cpu_usage to 1 and gpu_usage to 1 and mulitporcessor setting to 100%. This will make me run 15 cpu tasks and 1 gpu task. Oh, so that's why everyone sets cpu_usage to less than 1? What should I pick, 0.1, 0.01 or? |
juan BFP Send message Joined: 16 Mar 07 Posts: 9786 Credit: 572,710,851 RAC: 3,799 |
My "multiprocessors, use at most" setting was leaving 1 core free. But now I set cpu_usage to 1 and gpu_usage to 1 and mulitporcessor setting to 100%. This will make me run 15 cpu tasks and 1 gpu task. You use <cpu_usage>0.04 for MB, <cpu_usage>1.0 for AP. There are two diferent setting depending what type of WU you want to crunch. You could see that in the app_config file suplied. If you only crunch MB you could use only the MB part of the file and use only 0.04 and leave only one core free for all the 2 or 3 WU crunching in the GPU. Even in my hosts that not crunch AP i leave the file with both options enabled and choose what type of WU i want to crunch in the preferences page. |
Steve S Send message Joined: 19 Jan 13 Posts: 27 Credit: 6,274,024 RAC: 0 |
My "multiprocessors, use at most" setting was leaving 1 core free. But now I set cpu_usage to 1 and gpu_usage to 1 and mulitporcessor setting to 100%. This will make me run 15 cpu tasks and 1 gpu task. I don't know how to do that, documentation doesn't explain it. There's only one app name setiathome_v7, so i can't have 2 configs, unless there's an undocumented syntax. |
juan BFP Send message Joined: 16 Mar 07 Posts: 9786 Credit: 572,710,851 RAC: 3,799 |
I don't know how to do that, documentation doesn't explain it. There's only one app name setiathome_v7, so i can't have 2 configs, unless there's an undocumented syntax. The basic file is: <app_config> <app> <name>setiathome_v7</name> <gpu_versions> <gpu_usage>0.5</gpu_usage> <cpu_usage>0.06</cpu_usage> </gpu_versions> </app> <app> <name>setiathome_enhanced</name> <gpu_versions> <gpu_usage>0.5</gpu_usage> <cpu_usage>0.06</cpu_usage> </gpu_versions> </app> <app> <name>astropulse_v6</name> <gpu_versions> <gpu_usage>0.5</gpu_usage> <cpu_usage>1.00</cpu_usage> </gpu_versions> </app> </app_config> As you could see there are separete cpu_usage for each type of WU. This file allow you to run 2 WU at a time on the GPU, if you want to run 3 just change the 0.5 for 0.33. But remember very few GPU´s can run 3 AP faster than 2. 3 MB i belive your GPU can handle, more no. Of couse YMMV, so you need to test. BTW the diference of use 0.04 or 0.06 is meaningless in MB. |
Steve S Send message Joined: 19 Jan 13 Posts: 27 Credit: 6,274,024 RAC: 0 |
I don't know how to do that, documentation doesn't explain it. There's only one app name setiathome_v7, so i can't have 2 configs, unless there's an undocumented syntax. Yea, that's from the forum post, but the GPU CUDA tasks are named setiathome_v7. Anyway I can just specify the defaults in app_config.xml? I did try to delete the file, but somewhere it remembers my cpu/gpu_usage settings. This exercise appears to have given me no advantage in performance. Thanks for helping though. |
juan BFP Send message Joined: 16 Mar 07 Posts: 9786 Credit: 572,710,851 RAC: 3,799 |
Yea, that's from the forum post, but the GPU CUDA tasks are named setiathome_v7. astropulse_v6 uses the GPU too, that´s why you need to make the configuration for it too. If you dont run AP just leave aside and forget it. But not forget for now the AP crunching produces 2x credits than MB crunching. But that could change any time. To use the defaults (1 WU at a time on the GPU) you just need to delete the file and reload the boinc. To be totaly sure the file is working you must exit the boinc inself not only the manager, this is a common error and why the changes not work. If you use this file you will sure get a gain in performance don't give up. The GPU, specialy a 660TI running 2WU at a time, is able to produce almost the same RAC itself than the rest of your 16 processors CPU combined. Maybe someone else could help you with a better english. |
©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.