Message boards :
Number crunching :
v0.23 performance drop?
Message board moderation
Author | Message |
---|---|
Send message Joined: 28 Aug 07 Posts: 52 Credit: 8,353,747 RAC: 0 |
Tested on 6.10.21/.51/.52 Without any CPU job is running the performance is at 99% but with CPU the performance dropped to around 74-82%, collatz and dnetc are ok. Catalyst 10.4, Win7 64, HD5850 |
Send message Joined: 3 May 10 Posts: 14 Credit: 15,793,156 RAC: 0 |
I have the same system and get the same problem. I dont allow MilkyWay to do work on the CPU. If I dont have a BOINC task running on the CPU, I get a WU done in 110 sec. If I let the CPU do work (in my case Climate prediction), it goes up to 130 secs. What is also strange that sometimes I get down to 115 secs, but only when the PC is also in use (listening to music, browsing Internet etc). Any idea what is causing this? |
Send message Joined: 3 May 10 Posts: 14 Credit: 15,793,156 RAC: 0 |
I did some testing: I booted my PC and started BOINC immediately and got a WU completion time of 130 sec. After half an hour, I started used Windows Media Player to play some music and the completion time immediately dropped down to 115 sec. My system: Win7 Prof x64, i5-750, HD5850 What could be causing this strange behaviour? |
Send message Joined: 17 Feb 08 Posts: 363 Credit: 258,227,990 RAC: 0 |
Tested on 6.10.21/.51/.52 That seems to be a driver issue. I've experienced something similar with Cat 10.3 and 10.4 here. Reverting back to 10.2 solved the problem. Join Support science! Joinc Team BOINC United now! |
Send message Joined: 10 Oct 07 Posts: 79 Credit: 69,337,972 RAC: 0 |
hello, I'm not an expert but if you use your monitor for anything graphical or video or browser for sites with graphics or video on them like u-tube etc etc then your asking your card to do two jobs i get the same effect to the same sort of degree the thing that shows this effect up worst is playing a DVD where the app and the video fight so much the dvd litterally almost stops and starts and the speed of the work units plummets best regards Ian |
Send message Joined: 3 May 10 Posts: 14 Credit: 15,793,156 RAC: 0 |
But I have the complete opposite effect: If I dont use my PC at all, but crunch on the CPU, the work gets done the slowest. If it is in light use, but also crunching on the CPU, works gets done faster. And if I stop the crunching of the other BOINC task on my CPU, works gets done fastest. btw, I allowed crunching on the GPU while the computer is in use. |
Send message Joined: 14 Feb 09 Posts: 999 Credit: 74,932,619 RAC: 0 |
It almost sounds like the computer is throttling back the CPU when it thinks it does not need full power. Make certain that you have the energy settings to high performance. |
Send message Joined: 30 Apr 10 Posts: 10 Credit: 27,768,203 RAC: 0 |
I did some testing: turn off the visualization in windows media player, and you should see your performance go back up slightly, if it was on. Also - using any application while your computer is processing WU, (GPU, or CPU) will cause a slight dip in processing time. Even though what you seem to be doing may not give you the feeling you're taxing the computer, you actually are. |
Send message Joined: 31 Mar 10 Posts: 12 Credit: 13,722,511 RAC: 0 |
I've got an Core i7 920 System and noticed the same behaviour a while ago while trying to overclock my 5850. As far as i found out, the GPU Usage and therfore the performance drops dramatically if all Cores are at 100%. GPU usage never exceeded 82%. The solution was to configure Boinc to only use 90% of the Processors on Multicore Systems. This freed one Core and raised the Milkyway GPU usage to 99%. |
Send message Joined: 4 Oct 08 Posts: 1734 Credit: 64,228,409 RAC: 0 |
I have an HD5850 on a Core 2 quad (Penryn) as a QX9650. The 4 cores are running at 100%, on Einstein and FreeHAL (usually for the latter), and the GPU still runs at 95%+. I normally try and knock the GPU load back to 90% - 94% as the screen refresh becomes sluggish. I am not aware that I am increasing the crunch times of the Milkyway WUs. Go away, I was asleep |
Send message Joined: 30 Dec 07 Posts: 311 Credit: 149,490,184 RAC: 0 |
I think Nehalem class CPUs are possibly more prone to contention slowdowns than Penryns because of hyperthreading and less L2 cache. I remember when I ran Einstein on all 8 CPU cores my MilkyWay GPU times increased quite a lot. Since that time I have usually left a core free to support the GPU. Now that I have added a 5970 to the 5870 I leave two cores free. I don't turn them off in the BOINC preferences though, I assign them to the GPU in the app_info.xml file. So running 6 MilkyWay tasks on the 3 GPU cores I use avg_ncpus of 0.4. For running 3 Collatz tasks I use avg_ncpus of 0.8. Also max_ncpus of 2 for both projects. If the GPU project runs out of work while I am away from the computer or asleep the CPU project uses all 8 cores again. I don't lose that much on the CPU projects because many of them run individual tasks faster on only 6 cores rather than on all 8. As a plus the CPU stays cooler which is very useful in summer. When I run AQUA with this configuration it will still run on all 8 cores with the BOINC version 6.10.29 I use. |
Send message Joined: 3 May 10 Posts: 14 Credit: 15,793,156 RAC: 0 |
Finally somone with the same problem ... Thank you for your advice. Sadly, leaving a whole CPU core free for assiting the GPU is not an option for me, but as I said, leaving WMP running in the background running on repeat has almost solved the problem for me (just 5 secs slower). I guess its because WMP stops Climate Prediciton from using all the CPU time. |
©2024 Astroinformatics Group