| Index | Recent Threads | Unanswered Threads | Who's Active | Guidelines | Search |
 |
World Community Grid Forums
Category: Active Research
Forum: Mapping Cancer Markers Forum
Thread: Performance comparison of MCM on x86(32bit) Vs x64(64bit) and Windows Vs Linux |
| No member browsing this thread |
|
Thread Status: Active Total posts in this thread: 27
|
|
| Author |
|
|
widdershins
Veteran Cruncher Scotland Joined: Apr 30, 2007 Post Count: 674 Status: Offline Project Badges: 
|
Have you looked at WUProp for comparisons. That is kind of what they do... http://wuprop.boinc-af.org/ http://wuprop.boinc-af.org/results/delai.py Edit: http://wuprop.boinc-af.org/results/projet.py?...on=Mapping+Cancer+Markers The figures are interesting, but should be taken with a pinch of salt. There is no mention of sample size and some of the stats make no sense when compared with other processors in the same series unless other factors are in play on whatever machine produced the stats e.g. other stuff running as well or processor being throttled. A limited sample size for some processors may have led to some results being skewed by one or two inaccuracies due to the sort of reasons mentioned above. |
||
|
|
Jim1348
Veteran Cruncher USA Joined: Jul 13, 2009 Post Count: 1066 Status: Offline Project Badges: 
|
I am running MCM on a Haswell 4771 under Win7 64-bit (4 virtual cores), and also on a Haswell 4770 under Linux Mint 16 Cinnamon 64-bit (8 virtual cores). To my surprise, the Linux machine (a new build) is running more slowly than the Win7 machine, by over an hour per work unit. Here are some comparable work units:
Win7 64-bit: MCM1_0002216_7956_1 04:19:15 (04:19:02) 99.92 (CPU %) MCM1_0002216_8793_1 04:18:23 (04:18:09) 99.91 (CPU %) Linux 64-bit: MCM1_0002216_7967_2 05:45:41 (05:38:35) 97.95 (CPU %) MCM1_0002216_0608_1 05:44:23 (05:37:35) 98.03 (CPU %) The similar CPU percentages indicate to me that the number of cores (4 versus 8) is not affecting the results, and there is nothing else running (e.g., no GPUs) on the Linux machine. Now I am new to Linux, and it is possible that there is some way to tweak it up, or there is some hardware problem causing the 4770 to run more slowly. But my next step is to replace Linux with WinXP and see how that works. |
||
|
|
Ingleside
Veteran Cruncher Norway Joined: Nov 19, 2005 Post Count: 974 Status: Offline Project Badges: 
|
The similar CPU percentages indicate to me that the number of cores (4 versus 8) is not affecting the results, and there is nothing else running (e.g., no GPUs) on the Linux machine. Since both computers has only 4 real cores, you should expect major differences between running 4 and running 8 MCM-instances. If where hadn't been any benefit from HT you would have had atleast a doubling of cpu-time, but since HT can give maybe 20% benefit it's not doubled. Based on your numbers, the Windows-computer crunches 22.2 tasks/day while the Linux crunches 34.1 tasks/day, meaning 53 % more tasks/day. Even if the Windows-computer would maybe produce 27 tasks/day by running 8 task/cpu, the Linux-computer would still be 26 % faster.  "I make so many mistakes. But then just think of all the mistakes I don't make, although I might." |
||
|
|
Sgt.Joe
Ace Cruncher USA Joined: Jul 4, 2006 Post Count: 7699 Status: Offline Project Badges: 
|
I am running MCM on a Haswell 4771 under Win7 64-bit (4 virtual cores), and also on a Haswell 4770 under Linux Mint 16 Cinnamon 64-bit (8 virtual cores). To my surprise, the Linux machine (a new build) is running more slowly than the Win7 machine, by over an hour per work unit. Here are some comparable work units: Win7 64-bit: MCM1_0002216_7956_1 04:19:15 (04:19:02) 99.92 (CPU %) MCM1_0002216_8793_1 04:18:23 (04:18:09) 99.91 (CPU %) Linux 64-bit: MCM1_0002216_7967_2 05:45:41 (05:38:35) 97.95 (CPU %) MCM1_0002216_0608_1 05:44:23 (05:37:35) 98.03 (CPU %) The similar CPU percentages indicate to me that the number of cores (4 versus 8) is not affecting the results, and there is nothing else running (e.g., no GPUs) on the Linux machine. Now I am new to Linux, and it is possible that there is some way to tweak it up, or there is some hardware problem causing the 4770 to run more slowly. But my next step is to replace Linux with WinXP and see how that works. Be very careful about drawing conclusions from different WU in MCM1 even if they are from the same batch. Even in the same batch there is variability of run time between units done on the same machine. The only true comparison would be if you could catch the exact same WU on the two different machines - that is you would be your own wingman. (But I do not know if they mix wingmen between OS's or if the wingmen need to belong to the same OS.) Cheers
Sgt. Joe
*Minnesota Crunchers* |
||
|
|
Jim1348
Veteran Cruncher USA Joined: Jul 13, 2009 Post Count: 1066 Status: Offline Project Badges:
|
Since both computers has only 4 real cores, you should expect major differences between running 4 and running 8 MCM-instances. If where hadn't been any benefit from HT you would have had atleast a doubling of cpu-time, but since HT can give maybe 20% benefit it's not doubled. Good point about the benefit of HT. But rather than estimating it, I am going to put all 8 cores of the Win7 machine on MCM and let it run for a while. |
||
|
|
Jim1348
Veteran Cruncher USA Joined: Jul 13, 2009 Post Count: 1066 Status: Offline Project Badges:
|
Be very careful about drawing conclusions from different WU in MCM1 even if they are from the same batch. Even in the same batch there is variability of run time between units done on the same machine. The only true comparison would be if you could catch the exact same WU on the two different machines - that is you would be your own wingman. (But I do not know if they mix wingmen between OS's or if the wingmen need to belong to the same OS.) Cheers I see some batches with a lot of variation, but the one I chose seemed to be quite consistent on each machine, and will choose such batches for my next test. [Edit 2 times, last edit by Jim1348 at Feb 16, 2014 2:45:57 AM] |
||
|
|
Sgt.Joe
Ace Cruncher USA Joined: Jul 4, 2006 Post Count: 7699 Status: Offline Project Badges:
|
Be very careful about drawing conclusions from different WU in MCM1 even if they are from the same batch. Even in the same batch there is variability of run time between units done on the same machine. The only true comparison would be if you could catch the exact same WU on the two different machines - that is you would be your own wingman. (But I do not know if they mix wingmen between OS's or if the wingmen need to belong to the same OS.) Cheers I see some batches with a lot of variation, but the one I chose seemed to be quite consistent on each machine, and will choose such batches for my next test. Keep us informed. Thanks Cheers
Sgt. Joe
*Minnesota Crunchers* |
||
|
|
Jim1348
Veteran Cruncher USA Joined: Jul 13, 2009 Post Count: 1066 Status: Offline Project Badges:
|
I have the results for 8 virtual cores running on both machines, for the only series running thus far on both (2270). It is not as consistent as I would like, but seems to show similar performance for both machines. The Linux machine is dedicated and just sits there; I don't even do web browsing on it. The Win7 machine is my main PC, but I disabled Folding on the GPUs for this check, and minimized its other uses.
Linux Mint 64-bit (Haswell 4770): MCM1_0002270_3960_0 06:38:56 (06:37:40) 99.68 MCM1_0002270_3310_2 07:13:11 (07:11:44) 99.67 MCM1_0002270_6021_1 09:09:13 (09:07:32) 99.69 MCM1_0002270_5935_1 05:10:18 (05:09:17) 99.67 MCM1_0002270_5568_1 06:36:05 (06:34:54) 99.70 MCM1_0002270_4630_0 05:05:47 (05:04:49) 99.68 Average: 6.8 hours Windows 7 64-bit (Haswell 4771): MCM1_0002270_3532_0 05:34:13 (05:28:56) 98.42 MCM1_0002270_3955_1 05:25:59 (05:22:01) 98.78 MCM1_0002270_3966_0 06:06:49 (06:03:08) 99.00 MCM1_0002270_3606_1 07:42:28 (07:38:25) 99.12 MCM1_0002270_4620_0 08:01:54 (07:58:03) 99.20 MCM1_0002270_4388_1 06:53:04 (06:48:29) 98.89 MCM1_0002270_4703_1 04:11:51 (04:10:28) 99.45 MCM1_0002270_3345_1 05:54:38 (05:51:10) 99.02 Average: 6.2 hours The Haswell 4770 on the Linux machine has a base frequency of 3.4 GHz, whereas for the 4771 on Win7 it is 3.5 GHz, but they both Turbo-Boost to 3.9 GHz. I checked the Win7 machine, and it is running at 3.7 GHz, but don't have a way to check the Linux PC. However, it is running in the default "Ondemand" mode, so it should be about the same. If there are any significant differences that become apparent when I convert the 4770 machine to WinXP I will post back, but otherwise I am just assuming that it will be about the same as for the Win7 machine (the reason for XP has to do with the GPU project, not the CPU). |
||
|
|
Jim1348
Veteran Cruncher USA Joined: Jul 13, 2009 Post Count: 1066 Status: Offline Project Badges:
|
Win XP versus Win7 64-bit
Here are the results comparing Win XP (32-bit) on the Haswell 4770 to Win7 (64-bit) on the Haswell 4771. Both CPUs were running at 3.7 GHz as measured by Core Temp, and the memories were all at 1333 MHz (9-9-9-24), so hardware differences should not be a factor. The series 2372 was very consistent, and I needn't average them. The results show Win7 64-bit to be almost 13 percent faster than Win XP 32-bit. Whether than same difference applies with fewer than all 8 cores running is another question. WinXP 32-bit (Haswell 4770) MCM1_0002372_1702_0 04:18:55 (04:18:53) 99.99 MCM1_0002372_2613_0 04:18:19 (04:18:18) 99.99 MCM1_0002372_2037_1 04:18:33 (04:18:31) 99.99 MCM1_0002372_1786_0 04:18:18 (04:18:17) 99.99 MCM1_0002372_2600_0 04:18:09 (04:18:07) 99.99 MCM1_0002372_2597_0 04:17:53 (04:17:51) 99.99 MCM1_0002372_2473_0 04:18:03 (04:18:00) 99.98 MCM1_0002372_2608_0 04:18:03 (04:18:00) 99.98 MCM1_0002372_2612_1 04:19:31 (04:19:28) 99.98 MCM1_0002372_1512_1 04:17:52 (04:17:50) 99.99 Win7 64-bit (Haswell 4701) MCM1_0002372_5887_0 03:49:07 (03:48:43) 99.83 MCM1_0002372_6748_1 03:49:22 (03:48:59) 99.83 MCM1_0002372_5772_0 03:49:29 (03:48:36) 99.62 MCM1_0002372_7029_1 03:49:06 (03:48:36) 99.78 MCM1_0002372_7041_1 03:49:01 (03:48:32) 99.79 MCM1_0002372_6616_1 03:49:01 (03:48:35) 99.81 MCM1_0002372_7191_0 03:49:25 (03:48:30) 99.60 MCM1_0002372_5835_1 03:48:48 (03:48:21) 99.80 MCM1_0002372_6837_0 03:49:07 (03:48:36) 99.77 MCM1_0002372_5553_1 03:48:55 (03:48:21) 99.75 |
||
|
|
Randzo
Senior Cruncher Slovakia Joined: Jan 10, 2008 Post Count: 339 Status: Offline Project Badges:
|
Thank you Jim, interesting figures.
I may be due to Win7 better memory management. Have you been considering Linux? |
||
|
|
|