I recently upgraded from Win XP Pro 32bit to Win 7 Pro 64 bit SP1. My Dell Precision 390 with a GeForce 9600 GSO on PCI-E x16 is now chunking through Einstein work slower than before. I also upgraded to Nvidia driver 337.88 at the same time. Unfortunately I did not make careful note of conditions before the switch and so can't say for certain what specifically got slower. I only have 4 GB of RAM so I suspect that that is what is limiting performance. My only evidence of a slowdown is my daily output as recorded at http://boincstats.com/en/stats/-1/user/detail/110425/lastDays. Any advice or kibbitzing is welcome.
-C
Copyright © 2024 Einstein@Home. All rights reserved.
Slowdown after upgrade
)
WinXP is generally faster in GPU computing, because Vista and later implement the new WDDM architecture that takes additional performance penalty (overhead).
-----
RE: My only evidence of a
)
I would not trust credits as an accurate measure of crunching speed. The credits depend too much on your wingmen finishing their part.
A better measure is to look at the runtimes in your Tasks page here at E@H.
RE: I recently upgraded
)
Your 4gb of ram is also hurting you very badly. You are running gpu units and I have found on my own pc's, even XP, that that is just barely enough to run the units in a decent amount of time. For instance a pc with 16gb of ram is doing the units in 1 hour less time then one with 4gb of ram. I cannot upgrade my machine, you may not be able to either, so I had to stop all cpu crunching to make the units speed back up again. My XP machine is only a dual core machine and was only using one cpu to crunch cpu units, while the other core was free for the gpu to use. Units still had problems, but once I stopped the 2nd cpu core from also crunching they have mostly gone away and the gpu units are finishing faster.
One other thing is that the latest and greatest gpu software is not always better, gamers are one of their primary markets and they write their software for them, not us crunchers. If you like you can cruise thru the gpu software of the fastest crunchers and see the versions they use, if you see alot of them not using the latest version, you too might want to avoid it. Several versions of the Nvidia software actually slowed down our crunching!! Also there are usually a few threads at each project about the newest versions of the software, read them and see what others are saying before you update your own, essentially letting them be the guinea pigs.
RE: RE: A better measure
)
RE: RE: A better measure
)
When you posted your first message, I didn't have time to reply but I did have a quick look through your tasks list at the time and there was a single BRP5 GPU task dated around June 07 or 08 (if I remember correctly) that had taken around 36.5ksecs to crunch. I don't know if that date is prior to your changes or not. Recent tasks seem to be doing better than that, averaging around 34ksecs although there is a bit of variation.
Your CPU is showing as 1.86GHz and your GPU is a GeForce 9600GSO (384MB). I'm guessing the machine is a laptop based on the low CPU speed. If so, I guess you can't change the things that would really make a difference - GPU and CPU.
The 4GB RAM is unlikely to be a problem if you are just using the machine for 'office' type work, apart from crunching. I have lots of machines with 4GB RAM running just as fast as identical hardware with 8GB RAM. Linux has a nice graphical display of memory usage and my 4GB machines show more than 50% of physical RAM being free and zero swap space usage. The 8GB machines show 75% free physical memory. Those are quad core machines running 4 concurrent GPU tasks and 2 CPU tasks.
On your machine you are also running CPU tasks on both cores by the look of things. There seems to be quite a difference between CPU time and elapsed time (run time) for these tasks, indicating that GPU support duties are having quite an impact. It would be interesting to see what would happen if you allowed BOINC to use only one core for CPU crunching. You can do this through BOINC computing prefs (50% of processors) and I would imagine that with one free core, you might get a speedup of both the GPU task and the remaining CPU task. I think it would be worth trying that to see what happens. You can easily revert the change if you don't like the results :-).
If you're already running with 1 free core, then you are probably doing as well as you can and spending money on more memory is unlikely to make much of an improvement. I could easily be wrong as I'm basing that opinion on Linux performance which may be better at memory management than Windows.
Cheers,
Gary.
Per a suggestion via PM, I
)
Per a suggestion via PM, I reverted the Nvidia driver to version 280.26. GPU tasks are now running about 20% faster.