As for my intentions, I have 12 cores of 8th generation Intel i7, and integrated Intel GPU(UHD 630) and this beast of a Turing GPU so when I'm not compiling or gaming I'd like to use most of the cores and much of the GPU capacity to run cuda projects.
Currently:
1. about half of the Intel GPU is in use on a single work unit: einsteinbinary_BRP4
2. About 2% of the Nvidia GPU is in use on a single work unit hsgama_FGRPB1G
3. 12 iterations of einstein_010D1 are in memory but one appears idle (probably working on the GPU units)
If you could help me out with the app_confi, I'd be very appreciative.
It needs to be in the BOINC projects/einstein folder as its specific to Einstein. Make sure its file extension is .xml as notepad likes to tack a .txt on the end.
I suggest you stop using the intel GPU they slow the CPU down by about a factor of 5. By all means use it to display a screen but for compute tasks it slows the CPU down.
If those O1OD1 work units are running on the CPU what sort of times are you seeing for them? My i7-8700's are around 7 and a half hours. I don't have the K version of the CPU (tying to keep the heat and power bills down) so I expect yours will be quicker.
I've tried to get loads of Nvidia work units going but I can't seem to go above 8.
I currently have 1/8 set for GPU and 1 for CPU so that leaves 4 cores to just run the standard CPU app.
What's frustrating is the Nvidia card is only running at 10% but if I try to get more going, I get all kinds of computational errors and even one blue_screen. I haven't seen that in years.
What's frustrating is the Nvidia card is only running at 10% but if I try to get more going, I get all kinds of computational errors and even one blue_screen. I haven't seen that in years.
Yet more tasks on the Nvidia is not the direction to greater efficiency from where you are.
As it happens I am currently running an RTX 2080 on Einstein on the machine on which I am typing this note. For the current flavor of Einstein work, I see GPU load (as represented by GPU-Z) in the low 90s with a single Einstein Gamma-Ray Pulsar task running. Going to higher multiplicities gets a little more productivity at 2X than 1X, but little enough that I choose not to do it. Going yet higher gives infinitesimal further improvement. I'm pretty sure the improvement has turned negative by the time you get to 8X, though I've not actually tried that.
Better to pay attention to total throughput. How many GRP tasks of the current type are completed in a day? If you make an adjustment and that number gets better, you have an improvement. Nothing else matters.
If, truly, you somehow have Nvidia utilization down anywhere near 10% something in your system configuration is greatly robbing resources. The easiest one is the CPU support task. The current Einstein Windows GRP task uses a polling loop to alert the CPU to attend the GPU needs. So anything that means the CPU support task is not active at the moment the GPU needs help translates directly into increased elapsed time, reduced utilization, and lower total productivity. "Anything" in this case can mean BOINC CPU tasks, the CPU support task for other GPU tasks, and other work of any kind.
It puzzles me that your system logged a good many successful completions with elapsed times on the order of 600 seconds. That is reasonably healthy. Mine run about 520 seconds, but I am running some overclock on the 2080, and have stripped my system of most distractions, save for support of Einstein GRP running 1X on a 1060, so the 2080 is getting pretty full support.
There is no way your 2080 was running at 10% when it was returning Einstein GRP tasks with elapsed times of around 600 seconds on November 30 and December 1.
Oh, by the way, welcome to the Einstein Turing club.
As for my intentions, I have
)
As for my intentions, I have 12 cores of 8th generation Intel i7, and integrated Intel GPU(UHD 630) and this beast of a Turing GPU so when I'm not compiling or gaming I'd like to use most of the cores and much of the GPU capacity to run cuda projects.
Currently:
1. about half of the Intel GPU is in use on a single work unit: einsteinbinary_BRP4
2. About 2% of the Nvidia GPU is in use on a single work unit hsgama_FGRPB1G
3. 12 iterations of einstein_010D1 are in memory but one appears idle (probably working on the GPU units)
If you could help me out with the app_confi, I'd be very appreciative.
Regards,
Doug
Where did you put the
)
Where did you put the app_config?
It needs to be in the BOINC projects/einstein folder as its specific to Einstein. Make sure its file extension is .xml as notepad likes to tack a .txt on the end.
I suggest you stop using the intel GPU they slow the CPU down by about a factor of 5. By all means use it to display a screen but for compute tasks it slows the CPU down.
BOINC blog
C:\ProgramData\BOINC\proje
)
C:\ProgramData\BOINC\projects\einstein.phys.uwm.edu
Checking using the command line, just in case Windows Explorer is messing with my view...
BINGO... there was a .txt thrown on the end which wasn't showing up in Explorer.
What a rookie mistake.
I've been on Linux too long.
Checking if it works now...
The config file is working
)
The config file is working fine now.
Current status:
4 Intel GPU jobs
4 NVidia GPU jobs
10 Intel CPU jobs
So you think I should shut down the Intel GPU? Should I set the specific applications to 0?
Targeting 8 GPU work units
)
Targeting 8 GPU work units works for the Nvidia, but it mirrors 8 on the intel side as well.
How do I get them decoupled?
Thanks,
Doug
Correction, I only have 12
)
Correction, I only have 12 cores, so it's 6 GPU workunits apiece.
This leaves the Intel GPU hammered and the Nvidia still only doing about 9% utilization.
To disable the intel GPU I
)
To disable the intel GPU I would put a cc_config.xml into the boinc data folder and set the <ignore_intel_dev> flag.
See https://boinc.berkeley.edu/wiki/Client_configuration for details on what you can throw into the cc_config.
If those O1OD1 work units are running on the CPU what sort of times are you seeing for them? My i7-8700's are around 7 and a half hours. I don't have the K version of the CPU (tying to keep the heat and power bills down) so I expect yours will be quicker.
BOINC blog
Yes, I got the intel GPU
)
Yes, I got the intel GPU turned off.
I've tried to get loads of Nvidia work units going but I can't seem to go above 8.
I currently have 1/8 set for GPU and 1 for CPU so that leaves 4 cores to just run the standard CPU app.
What's frustrating is the Nvidia card is only running at 10% but if I try to get more going, I get all kinds of computational errors and even one blue_screen. I haven't seen that in years.
Puzzling... the Nvidia work
)
Puzzling... the Nvidia work units are processing about 1/6 the speed the did using the defaults.
I'm going to return to defaults and see if I can get back there.
I'm now seeing an hour processing time when I had seen 9 minutes. That was just looking at the Nvidia, but the Intel GPU was chugging along as well.
Dougga wrote:Yes, I got the
)
Good riddance
Yet more tasks on the Nvidia is not the direction to greater efficiency from where you are.
As it happens I am currently running an RTX 2080 on Einstein on the machine on which I am typing this note. For the current flavor of Einstein work, I see GPU load (as represented by GPU-Z) in the low 90s with a single Einstein Gamma-Ray Pulsar task running. Going to higher multiplicities gets a little more productivity at 2X than 1X, but little enough that I choose not to do it. Going yet higher gives infinitesimal further improvement. I'm pretty sure the improvement has turned negative by the time you get to 8X, though I've not actually tried that.
Better to pay attention to total throughput. How many GRP tasks of the current type are completed in a day? If you make an adjustment and that number gets better, you have an improvement. Nothing else matters.
If, truly, you somehow have Nvidia utilization down anywhere near 10% something in your system configuration is greatly robbing resources. The easiest one is the CPU support task. The current Einstein Windows GRP task uses a polling loop to alert the CPU to attend the GPU needs. So anything that means the CPU support task is not active at the moment the GPU needs help translates directly into increased elapsed time, reduced utilization, and lower total productivity. "Anything" in this case can mean BOINC CPU tasks, the CPU support task for other GPU tasks, and other work of any kind.
It puzzles me that your system logged a good many successful completions with elapsed times on the order of 600 seconds. That is reasonably healthy. Mine run about 520 seconds, but I am running some overclock on the 2080, and have stripped my system of most distractions, save for support of Einstein GRP running 1X on a 1060, so the 2080 is getting pretty full support.
There is no way your 2080 was running at 10% when it was returning Einstein GRP tasks with elapsed times of around 600 seconds on November 30 and December 1.
Oh, by the way, welcome to the Einstein Turing club.