Everyday is a schoolday in E@H - been using linux for years and did not know of taskset - thank you Jeroen, something else to tweak!
Here is a small script that I run on my linux systems. I set it to run every so often via cron.
regex="hsgamma|universe"
i=4
for p in `pgrep -f $regex`; do
taskset -cp $i $p
((i+=1))
done
The particular system I have it running on is running both FGRP4 and Universe@Home apps with a partial load of 4-threads. I set the tasks to run on the last two cores of the first CPU (4-5) and the first two cores (6-7) of the second CPU. I do this to keep the power draw and temps down and also because the CPUs will run at a higher frequency of 3.73 GHz with a maximum 2-core load. i7z is a good application for monitoring frequency, c-states, and temps on Intel i7 CPUs in linux and I use it to confirm that everything is running the way I want it to.
I added another 8GB RAM and the time required to process an FGRP4 dropped by almost half. No kidding. From something like 54,000 seconds per unit to ~29,000 seconds.
So it must have been all of that memory overhead, maybe slowing things down with a lot of swapping on the hard drive. Additionally, the CPU is now using dual-channel memory mode.
The remaining mystery is why the damned GeForce GTX 960 (on Linux) takes about 1/3 longer than my GeForce GTX 950 on Windows. I'm pretty sure neither is overclocked. The Windows box also has an i3 3.6GHz and 8GB RAM.
I added another 8GB RAM and the time required to process an FGRP4 dropped by almost half. No kidding. From something like 54,000 seconds per unit to ~29,000 seconds.
So it must have been all of that memory overhead, maybe slowing things down with a lot of swapping on the hard drive. Additionally, the CPU is now using dual-channel memory mode.
As @logforme noted ... was it not using dual channel before (or just 1x8GB before)? If not you could change it back to single-channel (either bios or move the memory to the wrong slot) and compare. That is likely to be a big effect.
Quote:
The remaining mystery is why the damned GeForce GTX 960 (on Linux) takes about 1/3 longer than my GeForce GTX 950 on Windows. I'm pretty sure neither is overclocked. The Windows box also has an i3 3.6GHz and 8GB RAM.
I had a quick look at your Linux host, i noticed you improvement on FGRPB1 (i think a very similar app to FGRP4) how many CPU / GPU tasks are you running on each?
I would suggest a little testing, cease CPU apps and compare just running GPU apps, i find the GPU apps on my linux host suffer dramatically if competing with any CPU tasks.
Yes, it was running 1x8GB before, but I didn't expect the dual-channel mode would have such a dramatic effect because numerous benchmarking sites I looked at showed data that seemed to trivialize the advantages. Of course, most applications and games don't use multithreading the way BOINC does.
I also suspect the raw increase in RAM had to do with it too. BOINC would not use all of the cores until I bumped up the memory allocation BOINC is allowed to use, so it was pretty close to maxing out the RAM before.
So somewhere between doubling the RAM and doubling the channels to the RAM helped the CPU tasks quite a lot.
However, the GPU still sucks. It cut the CPU time the GPU tasks are using, but the time for the GPU to complete tasks remains unchanged. Instead of 20% *faster* than my Windows GTX950, the Linux GTX 960 is about 20% *slower*. And I followed the instructions for installing CUDA to ensure Linux native drivers aren't interfering. I'm pretty sure neither card is overclocked.
The other funny thing is, the GTX 960 is the only GPU running on this machine, while the i3 Windows box is running its Intel integrated GPU and the GTX 950. The CPU tasks are moving glacially slow, but both GPUs are ticking along nicely.
Yes, it was running 1x8GB before, but I didn't expect the dual-channel mode would have such a dramatic effect ...
You might be interested in this message I posted back in Dec 2014. Nothing to do with GPU times - just a rather dramatic reduction in CPU crunch time when going from single channel to dual.
And this is good news because my motherboard not only supports dual-channel, but quad-channel. The bad news is I already have 16GB RAM and adding another 16 GB is a little ridiculous (a lot of performance benchmarking sites say you get rapidly diminishing returns beyond 16GB).
I don't know what you're using for system monitoring, but I use hwinfo64 and it can be very useful for diagnosing things or tuning. Temps at various locations, load, individual core frequencies, etc in realtime - pretty much everything you can monitor from any sensor in the system.
I don't know what you're using for system monitoring, but I use hwinfo64 and it can be very useful for diagnosing things or tuning. Temps at various locations, load, individual core frequencies, etc in realtime - pretty much everything you can monitor from any sensor in the system.
I can't recall what I was using right now, but I was monitoring individual core temps and the GPU temp because I was really paranoid about overclocking the i7-5820K. And I've never used a water cooler before, and wasn't terribly confident about the contact between it and the CPU.
But the overclock tipped the temp up a few degrees C (both before and after at 100% load), but still way within the CPU rated temperature tolerance.
The GPU is something like 39C. I'm considering overclocking that after I try adding one more stick of RAM (for 24GB, 3 channels to CPU). I checked the specs for the 950 and the 960, and confirmed neither were overclocked so it is really irritating that the 950 keeps outperforming the 960. Especially since it's in my son's gaming machine and he keeps turning it off because some damned fool (me) told him about energy conservation! Even more oddly, that machine is running the integrated Intel GPU as well as the GTX 950.
RE: Everyday is a schoolday
)
Here is a small script that I run on my linux systems. I set it to run every so often via cron.
regex="hsgamma|universe"
i=4
for p in `pgrep -f $regex`; do
taskset -cp $i $p
((i+=1))
done
The particular system I have it running on is running both FGRP4 and Universe@Home apps with a partial load of 4-threads. I set the tasks to run on the last two cores of the first CPU (4-5) and the first two cores (6-7) of the second CPU. I do this to keep the power draw and temps down and also because the CPUs will run at a higher frequency of 3.73 GHz with a maximum 2-core load. i7z is a good application for monitoring frequency, c-states, and temps on Intel i7 CPUs in linux and I use it to confirm that everything is running the way I want it to.
Ok guys, mystery solved, I
)
Ok guys, mystery solved, I think.
I added another 8GB RAM and the time required to process an FGRP4 dropped by almost half. No kidding. From something like 54,000 seconds per unit to ~29,000 seconds.
So it must have been all of that memory overhead, maybe slowing things down with a lot of swapping on the hard drive. Additionally, the CPU is now using dual-channel memory mode.
The remaining mystery is why the damned GeForce GTX 960 (on Linux) takes about 1/3 longer than my GeForce GTX 950 on Windows. I'm pretty sure neither is overclocked. The Windows box also has an i3 3.6GHz and 8GB RAM.
RE: Additionally, the CPU
)
By only using half of the memory bandwidth you seriously handicapped your CPU.
RE: I added another 8GB RAM
)
As @logforme noted ... was it not using dual channel before (or just 1x8GB before)? If not you could change it back to single-channel (either bios or move the memory to the wrong slot) and compare. That is likely to be a big effect.
I had a quick look at your Linux host, i noticed you improvement on FGRPB1 (i think a very similar app to FGRP4) how many CPU / GPU tasks are you running on each?
I would suggest a little testing, cease CPU apps and compare just running GPU apps, i find the GPU apps on my linux host suffer dramatically if competing with any CPU tasks.
Yes, it was running 1x8GB
)
Yes, it was running 1x8GB before, but I didn't expect the dual-channel mode would have such a dramatic effect because numerous benchmarking sites I looked at showed data that seemed to trivialize the advantages. Of course, most applications and games don't use multithreading the way BOINC does.
I also suspect the raw increase in RAM had to do with it too. BOINC would not use all of the cores until I bumped up the memory allocation BOINC is allowed to use, so it was pretty close to maxing out the RAM before.
So somewhere between doubling the RAM and doubling the channels to the RAM helped the CPU tasks quite a lot.
However, the GPU still sucks. It cut the CPU time the GPU tasks are using, but the time for the GPU to complete tasks remains unchanged. Instead of 20% *faster* than my Windows GTX950, the Linux GTX 960 is about 20% *slower*. And I followed the instructions for installing CUDA to ensure Linux native drivers aren't interfering. I'm pretty sure neither card is overclocked.
The other funny thing is, the GTX 960 is the only GPU running on this machine, while the i3 Windows box is running its Intel integrated GPU and the GTX 950. The CPU tasks are moving glacially slow, but both GPUs are ticking along nicely.
RE: Yes, it was running
)
You might be interested in this message I posted back in Dec 2014. Nothing to do with GPU times - just a rather dramatic reduction in CPU crunch time when going from single channel to dual.
Cheers,
Gary.
That's very
)
That's very interesting.
This is one of the articles that made me suspect the extra channels were a bit overrated http://www.pcworld.com/article/2982965/components/quad-channel-ram-vs-dual-channel-ram-the-shocking-truth-about-their-performance.html Though his first chart claims to show quad-channel is so much better than dual-channel, all the subsequent charts show little difference. But again, BOINC certainly uses multithreading and all of those CPU cores even if a lot of daily-drivers don't.
And this is good news because my motherboard not only supports dual-channel, but quad-channel. The bad news is I already have 16GB RAM and adding another 16 GB is a little ridiculous (a lot of performance benchmarking sites say you get rapidly diminishing returns beyond 16GB).
I don't know what you're
)
I don't know what you're using for system monitoring, but I use hwinfo64 and it can be very useful for diagnosing things or tuning. Temps at various locations, load, individual core frequencies, etc in realtime - pretty much everything you can monitor from any sensor in the system.
http://www.hwinfo.com/download.php
RE: I don't know what
)
I'm its author and some of the features I have tuned especially based on my experience with running such projects ;-)
-----
I can't recall what I was
)
I can't recall what I was using right now, but I was monitoring individual core temps and the GPU temp because I was really paranoid about overclocking the i7-5820K. And I've never used a water cooler before, and wasn't terribly confident about the contact between it and the CPU.
But the overclock tipped the temp up a few degrees C (both before and after at 100% load), but still way within the CPU rated temperature tolerance.
The GPU is something like 39C. I'm considering overclocking that after I try adding one more stick of RAM (for 24GB, 3 channels to CPU). I checked the specs for the 950 and the 960, and confirmed neither were overclocked so it is really irritating that the 950 keeps outperforming the 960. Especially since it's in my son's gaming machine and he keeps turning it off because some damned fool (me) told him about energy conservation! Even more oddly, that machine is running the integrated Intel GPU as well as the GTX 950.