Message boards :
Number crunching :
How are people acheiving the super high RACs?
Message board moderation
Previous · 1 · 2
Author | Message |
---|---|
![]() ![]() Send message Joined: 20 May 99 Posts: 67 Credit: 109,352,237 RAC: 1 ![]() |
@ Grant: I'm running these servers exclusively for SETI number crunching. They are mostly unused servers and some older admin workstations in my lab. |
![]() ![]() Send message Joined: 4 Oct 00 Posts: 9541 Credit: 50,759,529 RAC: 60 ![]() ![]() |
Unused server? considered donating it to seti@home? ![]() In a rich man's house there is no place to spit but his face. Diogenes Of Sinope |
Cruncher-American ![]() ![]() ![]() Send message Joined: 25 Mar 02 Posts: 1513 Credit: 370,893,186 RAC: 340 ![]() ![]() |
How are people achieving the super high RACs? Obvious - by having sex with staff <g> |
-BeNt- ![]() Send message Joined: 17 Oct 99 Posts: 1234 Credit: 10,116,112 RAC: 0 ![]() |
How are people achieving the super high RACs? Now that made me chuckle. lol. Traveling through space at ~67,000mph! |
Darren Wright Send message Joined: 15 Jan 00 Posts: 92 Credit: 17,556,032 RAC: 0 ![]() |
I'm running Dell big-iron servers (R900, R910) with massive cooling and baseboard management...also power pigs btw ;) Super VM boxes, not so efficient SETI crunchers as far as RACs per THERM. a few Nvidia GPU's may make you shut some of them down. |
Grant (SSSF) Send message Joined: 19 Aug 99 Posts: 13882 Credit: 208,696,464 RAC: 304 ![]() ![]() |
@ Grant: Then the RAC for the 60 processor system is way down. My E6600 (2.4GHz C2D) running an optimised application had an RAC of around 2,400. So with an optimised application you should get at least 19,000 for each 16 processor machine (actually more due to the higher clock speed). The stock application is about a third of the speed of the optimised, so those machines should have an RAC of around 6,300 (which a couple do, the others are low- only recently started using them?). With the 60 core machine you should get an RAC of 65,000-70,000 with an optimised application, so around 22,000 running stock. It's RAC is only 13,400. So either it's only been in use for a short while, it spends a lot of time running other processes, or not all of the Cores are being utilised. Before buying & installing CUDA cards, i'd suggest using the Lunatics installer to install the best optimised application for each system and let them run those for a while & see how they handle it- they make the CPUs work *much* harder than the stock application. If after a few weeks the systems are still stable & returning valid work, then add a CUDA card or 2 to a system- making sure the Power supply is up to the job. And watch the RAC really climb. It generally takes 6-8 weeks before RAC will stabliise around it's new level after making changes. I haven't had a close look at the machines or the numbers, but with optimised applications & a CUDA card or 2 in each of them an RAC of over 1 million would look about right. Grant Darwin NT |
![]() ![]() Send message Joined: 27 Apr 11 Posts: 1932 Credit: 17,952,639 RAC: 0 ![]() |
|
![]() ![]() Send message Joined: 18 May 99 Posts: 6497 Credit: 34,134,168 RAC: 0 ![]() |
I'm running Dell big-iron servers (R900, R910) with massive cooling and baseboard management...also power pigs btw ;) To get an idea what will happen.. look at the number of "shaders" some of the video cards have. Each one is a "core" effectively. If someone pops in a video card with 512 shaders or more, The 16 core machines kind of pale in comparison. Rule 1: More cores/shaders takes you to the end of a mile faster. Rule 2: A mile varies. Janice |
![]() ![]() Send message Joined: 27 Apr 11 Posts: 1932 Credit: 17,952,639 RAC: 0 ![]() |
|
![]() ![]() Send message Joined: 20 May 99 Posts: 67 Credit: 109,352,237 RAC: 1 ![]() |
Ok, I took one of the 16 core machines and put in two nVidia older GPU cards I had collecting dust (2x GeForce 9400GT). I also installed the latest Lunatics app (v0.38). Now, is there any easy way to see a difference in performance? Or do I just need to sit back and wait for the RAC to climb? |
tbret ![]() Send message Joined: 28 May 99 Posts: 3380 Credit: 296,162,071 RAC: 40 ![]() ![]() |
Ok, I took one of the 16 core machines and put in two nVidia older GPU cards I had collecting dust (2x GeForce 9400GT). I also installed the latest Lunatics app (v0.38). Now, is there any easy way to see a difference in performance? Or do I just need to sit back and wait for the RAC to climb? You should very soon see CUDA work units that you would not otherwise have crunched, so you can look at those as "instant increase." Whatever advantage you bought with v0.38 you will first be able to see by comparing times for work units in a before / after way posted in your "valid" list of completed tasks. But for an idea of your real increase in RAC, you'll just have to wait a couple of weeks (or three, or a month) for it to stabilize at a new, higher, number. That assumes that the project is mostly up during that time. |
Grant (SSSF) Send message Joined: 19 Aug 99 Posts: 13882 Credit: 208,696,464 RAC: 304 ![]() ![]() |
Ok, I took one of the 16 core machines and put in two nVidia older GPU cards I had collecting dust (2x GeForce 9400GT). I also installed the latest Lunatics app (v0.38). Now, is there any easy way to see a difference in performance? Or do I just need to sit back and wait for the RAC to climb? Within a week you should see a significant increase in your RAC, however it will take 6-8 weeks before it gets to it's new range of operation. It would be worth checking your account page for the tasks on that machine, and keeping an eye on the number of Invalid or Error Work Units. Ideally both should be zero, but with the occasional outage & slow downs we've been having you may get a few Work Units that result in an Error, mostly due to download problems. Grant Darwin NT |
![]() ![]() Send message Joined: 4 Oct 00 Posts: 9541 Credit: 50,759,529 RAC: 60 ![]() ![]() |
you could also take a peek at the BOINCstats page. It shows your daily production which should increase significantly as your computer eats up the WUs. You still have to wait on wingmen so DOn't be surprised if the Credits don't blast off right away ![]() In a rich man's house there is no place to spit but his face. Diogenes Of Sinope |
![]() ![]() Send message Joined: 20 Aug 02 Posts: 3377 Credit: 20,676,751 RAC: 0 ![]() |
A quick look at your tasks list shows you have completed two GPU tasks. I took one and dug around to find one on your CPU with a comparable angle range and came up with these figures... on your CPU a 0.37AR WU took you a run time of 10,082 seconds, a CPU time of 10,021 seconds and gave you 93.49 credits. On your GPU it gave a run time of 8465 seconds and a CPU time (the time it took to load your GPU and do some other stuff) of 200.21 seconds and gave you 144.79 credits. With the new credit scheme they have come up with the credits are not that great a way to tell how they are doing but those ~1600 seconds will make a nice difference plus you have added two more cores to the cause. :-) Not bad for a couple of cards that were laying around collecting dust. ![]() PROUD MEMBER OF Team Starfire World BOINC |
![]() ![]() Send message Joined: 20 May 99 Posts: 67 Credit: 109,352,237 RAC: 1 ![]() |
So did I read correctly that even if I am not running any GPUs I should still install the Lunatics package because it is optimised for number crunching? |
![]() ![]() Send message Joined: 7 Jan 02 Posts: 497 Credit: 14,261,068 RAC: 67 ![]() ![]() |
I'm running Dell big-iron servers... I'll go ahead and assume it's not your power tab or personal hardware. If thus, be sure you don't NEZ yourself. |
![]() ![]() Send message Joined: 20 Aug 02 Posts: 3377 Credit: 20,676,751 RAC: 0 ![]() |
So did I read correctly that even if I am not running any GPUs I should still install the Lunatics package because it is optimised for number crunching? Oh yes definitely. It will make a world of difference. The stock apps are a one size fits all affair. The optimized apps allow you to take advantage of the instructions set specific to your machine for much better performance. ![]() PROUD MEMBER OF Team Starfire World BOINC |
![]() ![]() Send message Joined: 11 Sep 99 Posts: 6534 Credit: 196,805,888 RAC: 57 ![]() ![]() |
So did I read correctly that even if I am not running any GPUs I should still install the Lunatics package because it is optimised for number crunching? Yes SETI@home classic workunits: 93,865 CPU time: 863,447 hours ![]() |
Josef W. Segur Send message Joined: 30 Oct 99 Posts: 4504 Credit: 1,414,761 RAC: 0 ![]() |
So did I read correctly that even if I am not running any GPUs I should still install the Lunatics package because it is optimised for number crunching? Definitely so, if you continue to be interested enough to keep track of future changes. For that XEON X5570 system, you chose the SSE3 CPU version but the SSSE3X version would perform better. It would also run the SSE4.1 version but unless there's very fast RAM handled exquisitely by QPI the SSSE3X is probably better. Either one is significantly faster than the SSE3 version, but even the SSE3 version is likely to be twice as fast as stock 6.03. From the first 2 validations of GPU work, the "Average processing rate" is estimated at ~19.6, compared to ~12 for CPU tasks. The rate for CPU tasks is probably growing as the Lunatics SSE3 build works on those tasks originally sent to be done by stock 6.03. As a rough approximation, I expect the 2 GPUs will be about like adding 3 additional CPU cores to the system since the 9400 GT isn't a top-end crunching GPU. Joe |
Grant (SSSF) Send message Joined: 19 Aug 99 Posts: 13882 Credit: 208,696,464 RAC: 304 ![]() ![]() |
So did I read correctly that even if I am not running any GPUs I should still install the Lunatics package because it is optimised for number crunching? The SSSE3 version trippled my CPU output compared to the stock version. EDIT- that was back then the stock version didn't make as much use of optimisations as the current one does. Grant Darwin NT |
©2025 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.