can I squeeze more out of these cores?
14 Oct 2012 12:16:09 UTC
Topic 196569
(moderation:
I've got an 8 core processor of which I use 6 for crunching Einstein. One task per core. No gpu processing. I've noticed that I'm losing position in the stats, so that begs the question, is there a way I can do more with these 6 cores than just running one task per core? SIV64 says there's 98-100% utilization on those cores, so I'm wondering if it could handle possibly 2 tasks per core. Anybody got any thoughts on that?
Thanks,
mdawson
can I squeeze more out of these cores?
)
Unlike the situation on GPU cards where the system allows you to run (often but not always usefully) more than one task simultaneously on the GPU, for CPU tasks the system only allows you to run one task per virtual core.
If you have a physical 8 core CPU which is capable of being run hyperthreaded, then once you enable the hyperthreading (typically this is a BIOS setting) than from the boinc point of view you will be seen as having a 16 core CPU, and it would allow you to run 12 instead of 8 if you liked. Whether this is helpful or not depends on details of the application and of the system, but generally it does lead to modestly to moderately more output.
As to the subject of losing position, I suspect it is likely that most of the systems moving past you in the ranking order contain a GPU. At the moment on Einstein (and elsewhere where GPU's are supported) the computational performance per unit capital dollar or per unit electric power expenditure of a modern GPU is considerably higher than that of any CPU-based system.
[Edit: after typing that, I noticed that your system is an i7 960. That is a four core CPU which is being detected as an 8 processor system because it is already being run hyperthreaded--so you are already running more than one task per core]
On the software side, be sure
)
On the software side, be sure that you are not allowing the BRP apps to run on CPU, they take a really long time to be crunched there and they wont give you a fair amount of credits for the time used.
Another thing, do a comparison on the credits per time you get with the FermiLab and the Gravitational wave search... Ussually on intel CPUs the GWS gives a bit more credits per hour than the other, so you may opt out on one of these to get a better RAC... (you can choose all this things in the Einstein prefferences page)
On the harware side the only things you can do is to allow Einstein to use the 8 cores and to apply some Overclocking... But, then you may have responsiveness and heat issues...
The apps that run on CPU do it with the lowest possible priority, so its supossed that they get out of the way as fast as the host needs the CPU for anythig else, but, anyway it takes some time for the system to switch to another app and depending on your system usage it may lead to a certain degree of unresposiveness, and also as the CPU will be more used, then it will rise their temp... If it is worth to use them all is something you will need to test on your host...
Overclocking can also give you some improvement on performance, but ussually the gain is so small that its not worth the effort you will need to find the right amount of OC that your system can handle in a safe way. (bear in mind that just one invalid result due to overclock might be a greather loss on credits than the overall improvement due to OC)
But the truth is that no matter what, your relative position over the stats will keep falling if you dont use a GPU... I have an I7-2600 with a GT430 GPU and that cheap GPU gives me around the same credits per day than the 8 cores of the i7...
Horatio and Archae86, I do
)
Horatio and Archae86,
I do have a killer gpu and I use it for Collatz. I had experimented with E@H on gpu and cpu, but it was taking too many resources and was locking up my machine at the time. It was a dual core and I didn't know how to manipulate cpu affinity at the time, so I started Collatz.
On a side note, I never had much luck OC'ing, so I run at stock speeds. I do have a wicked cpu cooler though. When I installed it, temps dropped pert near 35-40 degrees C. per core. I had a lot of dust built up deep in the radiator down by the heat transfer plates. My new cooler is water cooled with double fans in series that attach on the case. I just wish the gpu cooled as nicely, but the onboard fan is doing ok. It runs at about 85% speed and 70c with some noise when Collatz CUDA40 app is running. When Collatz CUDA 31 is running, temp drops to about 61 C, but so does gpu utilization 24% or so vs. 90% or so.
I've got a GTX 260 which won't fit in to my case (my GTX 680 replaced it), and I've got an older geforce 8600GT that has a dead fan. I could use it if only I could source the fan. Fry's doesn't have one that fits.
I'm not really in this for the credits. It's all about the science for me. I dropped out for a few years and have not been able to reclaim my former position. I'd opt for more S6's', but as it is, I only get about 1 in 25 WU's that are BRPS. The rest are S6's. Somebody has to do the BRPS', right? Yes, the BRPS' take about 12.5 hrs cpu only, and the S6's at about 6 hrs., but I'm not dissatisfied. My system seems to be running as efficiently as possible in terms of crunching two different projects. I've got good numbers for both projects, but that is probably mostly due to longetivity in the projects rather than computational power which was only increased dramatically during the last year or so.
Thanks for the responses guys. I will just keep the status quo for now unless I decide to rejoin SETI.
mdawson
RE: I'm not really in this
)
The point is that the BRPs can be processed more efficiently on GPUs and on average they take between 30 mins to 1.5 Hour depending on the model of the GPU. In Einstein BPRs are the only GPU work, so for sure there are a lot of crunchers taking care of them...
Im just saying that your contribution and your electricity usage will be more efficient avoiding them on CPU and as a side effect you will get a better RAC, but of course, that doesnt means that you shouldnt do them on CPU...
By the way, now that you have more available cores on the CPU, and considering that current apps for BRP GPU are more optimized than before, may be you can try them again... I guess they will not clog your current system as you noticed before...
In my main Einstein cruncher wich is also my main computer for work, I have 2 560Tis running 2 concurrent BRP each, plus 6 cores doing S6 tasks and If not for the sound of the fans I barely notice they are running even when Im using the Process Lasso utility to rise the priority of the GPU apps...
(Messing with the affinity ussually doesnt help at all, indeed, it ussually interferes with the OS schedulling mechanism making everything worst...)