Message boards :
Number crunching :
Hyperthreading yes or no
Message board moderation
Author | Message |
---|---|
Bearcat Send message Joined: 10 Sep 99 Posts: 106 Credit: 10,778,506 RAC: 0 |
I just ran across the problem with multiple wu's on one cpu in another thread, which made me think about my i7 setup. Someone in that thread made the remark that even cpu's which are set up for hyperthreading have a small penalty, which makes total sense because of the switching back and forth between the threads on the same core. The value mentioned there was 15%, which I cannot confirm, but it sounds reasonable. I am currently using 7 of the theoretical 8 cores on my i7 for crunching plus a gtx275. Can someone please remind me why I did that ? And no, I am for the moment not concerned about the weekly wu limitations. I think I came to 7 out of 8 cores to have enough capacity left over to service the gpu, but not sure. |
soft^spirit Send message Joined: 18 May 99 Posts: 6497 Credit: 34,134,168 RAC: 0 |
I can only guess as to why, and can not give an honest answer whether it was good thinking or not.. But I am guessing you did not use the last core to: 1:give the GPU a core to drive it and 2: you wanted the computer to actually work like a computer too. Janice |
hiamps Send message Joined: 23 May 99 Posts: 4292 Credit: 72,971,319 RAC: 0 |
The highest my RAC ever got was with HT turned off and using 3 of 4 cores...I think mark sattler suggested it can't remember. Official Abuser of Boinc Buttons... And no good credit hound! |
Bearcat Send message Joined: 10 Sep 99 Posts: 106 Credit: 10,778,506 RAC: 0 |
The highest my RAC ever got was with HT turned off and using 3 of 4 cores...I think mark sattler suggested it can't remember. Hmmm. Lets just assume for a moment that a real core has a theoretical value of 100. Two virtual cores would then have 85 (100-15 penalty) or 42.5 each. If we compare 3 real cores to 7 virtual cores it would be like 300 compared to 297.5 value, and that looks like a win for 3 over 7, but it all hinges on the assumption of 15% penalty. If for example the wasted time is only 10% the 7 cores win clear over the 3 with 300 to 315. Has anyone ever actually tested that ? I realize that we are talking about 1-5% preformance here, but hey :) |
hbomber Send message Joined: 2 May 01 Posts: 437 Credit: 50,852,854 RAC: 0 |
Overall boost of using HT is between 30 and 50% depending how fast and with low latency your memory is. When HT is on, having fast and responsive memory is the key. U may have observed already, completions times of a single unit are longer, but your output is twice more units, netting u above numbers. When I was running CPU work primarily, that was my observation and I did tests several times, when someone raises this question over and over, and gets the answer to turn HT off. U don't need to dedicate core to GPU too. I was running in same way like u - overclocked i7-920 + GTX 275. GPU crunch times did not show any deviations, compared to same cards, on machines, which can provide all available CPU speed to serve GPU tasks. Your PC will be hardly responsive sometimes, not bcs your cores are loaded and running with SETI, but bcs GPU is loaded with CUDA. |
James Sotherden Send message Joined: 16 May 99 Posts: 10436 Credit: 110,373,059 RAC: 54 |
I run my i7 with HT on and all 8 cores crunch. The GPU a GTS 250 runs also. No over clock just running App Opps. I only have a problem when i get a VLAR On my GPU and I try and play solitare or watch a video. Then I gety the jerky movements. I dont know what the rac would peak out to, beacuase Seti allways went down before I could see it peak. I think it would be well over 8K. [/quote] Old James |
DJStarfox Send message Joined: 23 May 01 Posts: 1066 Credit: 1,226,053 RAC: 2 |
Hmmm. Lets just assume for a moment that a real core has a theoretical value of 100. Two virtual cores would then have 85 (100-15 penalty) or 42.5 each. If we compare 3 real cores to 7 virtual cores it would be like 300 compared to 297.5 value, and that looks like a win for 3 over 7, but it all hinges on the assumption of 15% penalty. If for example the wasted time is only 10% the 7 cores win clear over the 3 with 300 to 315. Has anyone ever actually tested that ? That's not how the math works. With HT on, each virtual core is 65% of a real core (because of x86 architecture, pipeline stalls, etc.). So, 7 virtual = 455 vs 3 real = 300. However, with the computer otherwise idle, the 7th core will actually run as fast as a real core. Therefore, we have 6 virtual + 1 real = 6 * 65 + 100 = 490. So, you get an extra 23% more performance using 7 virtual instead of 4 real cores. If he's actually doing work with one core, then the performance will be about 455 vs 300 with ht off (52% speedup having HT on with this scenario). So in conclusion, if you do work while BOINC is running on your computer, it would make perfect sense to keep HT on and limit to 7 cores out of 8. :) PS I always thought BOINC should have a compute setting for: % of processors to use when idle % of processors to use when user is busy (instead of all or nothing, as it is now) |
Todd Madson Send message Joined: 4 Aug 99 Posts: 71 Credit: 30,888,293 RAC: 15 |
Remember that the hyperthreading of today is much more efficient than the hyperthreading of the Pentium 4 days. The Corei7 seems to do a remarkable job of "acting like a hardware cpu" in the hyperthreading mode. On my system all I have eight workunits being crunched at once. No percievable system slowdowns. The biggest leap in performance I've found is when I went from 4 gigs of ram to 12 and that really increased my performance since the ram I had before was fast and so is this. |
Cruncher-American Send message Joined: 25 Mar 02 Posts: 1513 Credit: 370,893,186 RAC: 340 |
Remember that the hyperthreading of today is much more efficient than the hyperthreading of the Pentium 4 days. That memory improvement is surprising to me, as my 4GB 4 (real) core machine (Phenom II X4) + 3 x GT240 shows physical memory use of barely 2.1 GB in Task Manager (Vista Ultimate x64), so adding memory wouldn't seem to help.... and adding 8GB would seem to be excessive (remember, that usage also includes the OS, which is not trivial (thanks, MS). |
TheFreshPrince a.k.a. BlueTooth76 Send message Joined: 4 Jun 99 Posts: 210 Credit: 10,315,944 RAC: 0 |
Remember that the hyperthreading of today is much more efficient than the hyperthreading of the Pentium 4 days. If it is a Core i7 in most cases the mainboard supports Tripple Channel. 4Gb of memory means that it wasn't installed Tripple Channel. So by changing the memory to 3 modules (Tripple Channel) you have a wider memory bandwidth available for the CPU. And that increases performance. I now have an AMD 6 core CPU with DDR3 Double Channel support on the mainboard. The standard settings are "Unganged Mode" for the memory. "Unganged Mode" means that (2 or) 3 cores share the bandwith of one module. In most cases this is faster, as you can read on many hardware forums. But in the case of crunching for Seti I found out it is about 8% faster when I crunch in "Ganged Mode" (Double Channel support on). Rig name: "x6Crunchy" OS: Win 7 x64 MB: Asus M4N98TD EVO CPU: AMD X6 1055T 2.8(1,2v) GPU: 2x Asus GTX560ti Member of: Dutch Power Cows |
hiamps Send message Joined: 23 May 99 Posts: 4292 Credit: 72,971,319 RAC: 0 |
Remember that the hyperthreading of today is much more efficient than the hyperthreading of the Pentium 4 days. I was wondering about that also, I was just wondering why he would have 4 gigs? Official Abuser of Boinc Buttons... And no good credit hound! |
hiamps Send message Joined: 23 May 99 Posts: 4292 Credit: 72,971,319 RAC: 0 |
With all the server problems it is hard to tell but the highest my RAC ever got using HT was 37,000 without and using 3 cores I got to 52,000. EDIT: My machine runs a lot hotter using HT. Official Abuser of Boinc Buttons... And no good credit hound! |
TheFreshPrince a.k.a. BlueTooth76 Send message Joined: 4 Jun 99 Posts: 210 Credit: 10,315,944 RAC: 0 |
Remember that the hyperthreading of today is much more efficient than the hyperthreading of the Pentium 4 days. I tend to do the same sometimes when I'm low on budget :P I buy a high quality mainboard and a cheap CPU and RAM. When I raise a little funds, I upgrade the CPU and/or the RAM. Rig name: "x6Crunchy" OS: Win 7 x64 MB: Asus M4N98TD EVO CPU: AMD X6 1055T 2.8(1,2v) GPU: 2x Asus GTX560ti Member of: Dutch Power Cows |
hiamps Send message Joined: 23 May 99 Posts: 4292 Credit: 72,971,319 RAC: 0 |
Remember that the hyperthreading of today is much more efficient than the hyperthreading of the Pentium 4 days. I buy cheap and Overclock it as far as it will go....LOL Official Abuser of Boinc Buttons... And no good credit hound! |
TheFreshPrince a.k.a. BlueTooth76 Send message Joined: 4 Jun 99 Posts: 210 Credit: 10,315,944 RAC: 0 |
Remember that the hyperthreading of today is much more efficient than the hyperthreading of the Pentium 4 days. Always push it to the limits ;) Rig name: "x6Crunchy" OS: Win 7 x64 MB: Asus M4N98TD EVO CPU: AMD X6 1055T 2.8(1,2v) GPU: 2x Asus GTX560ti Member of: Dutch Power Cows |
Cruncher-American Send message Joined: 25 Mar 02 Posts: 1513 Credit: 370,893,186 RAC: 340 |
Remember that the hyperthreading of today is much more efficient than the hyperthreading of the Pentium 4 days. He (thst's ME) would have 4GB because, given the memory usage I see (as stated above) any more memory wouldn't do me any good. (I originally had 8GB - dual channel, of course) on both of my machines, and backed them down to 4GB when I saw what the actual usage was). |
hiamps Send message Joined: 23 May 99 Posts: 4292 Credit: 72,971,319 RAC: 0 |
Remember that the hyperthreading of today is much more efficient than the hyperthreading of the Pentium 4 days. I was under the impression that tripple channel would be better so I only have 3 gigs... Official Abuser of Boinc Buttons... And no good credit hound! |
Cruncher-American Send message Joined: 25 Mar 02 Posts: 1513 Credit: 370,893,186 RAC: 340 |
Triple channel is higher b/w than dual, so IS better, but my machines are AMD, and don't support it. |
hiamps Send message Joined: 23 May 99 Posts: 4292 Credit: 72,971,319 RAC: 0 |
Triple channel is higher b/w than dual, so IS better, but my machines are AMD, and don't support it. Well I learned something new today. Official Abuser of Boinc Buttons... And no good credit hound! |
Tim Lee Send message Joined: 15 Feb 00 Posts: 22 Credit: 32,655,046 RAC: 32 |
The actual performance is going to be very much subject to the application. I did an experiment when I first bought an i7 running my own multi threaded prime number generator. With this very simple code more than 4 threads never achieved any improvement. What was interesting was that performance did not begin to deteriorate until about 12 threads. However, with a more complex task, I would expect that the issues of cache stalls etc would mean that more than 8 threads might be optimal. I've been running Boinc off a usb flash memory stick for some years in order to allow the hard disks to power down when idle. Yes they do wear out after several years, but a flash memory from years ago is small enough that I usually have a few I no longer use ready to replace it. |
©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.