The recommendation is designed to be a good starting point but it may not suit all the different possible hardware combinations.
I agree to that. Especially in the case of laptops! In my case I did not only free one core, but disabled CPU work in general. Otherwise I would have serious temperature issues. Running my NVIDIA 610M and HD4000 is keeping it warm enough already.
The only way to achieve the best efficiency for a particular hardware combination is to experiment.
i´d suggest that as the First Law of Crunching?
There are a lot of hosts to compare against (especially your wing men), and the log file job_log_einstein.phys.uwm.edu.txt contain a history of jobs over time, so you can see trends / changes.
Errors and invalids should be zero on an efficient system.
Rather then get into a long drawn out explanation of this or that..let me just say 'personal experience' and let it go at that, okay? As a slight testament to back that up I will let my signature come thru this time:
Rather then get into a long drawn out explanation of this or that..let me just say 'personal experience' and let it go at that, okay? As a slight testament to back that up I will let my signature come thru this time:
Even that doesn't stop you learning something new each day. On this project, Gary knows what he's talking about.
I realize that the GPU has to get fed with data through the CPU, I just did not expect that a process using about 2% cpu power to not work correctly if it doesn't have a core of its own.... now, I'm wondering if it's more power efficient to run GPU tasks and have 2 of 8 cores more or less idle or run on all 8 cores and do no GPU work.
Rather then get into a long drawn out explanation of this or that..let me just say 'personal experience' and let it go at that, okay? As a slight testament to back that up I will let my signature come thru this time:
Even that doesn't stop you learning something new each day. On this project, Gary knows what he's talking about.
And from some folks, including Gary, I do learn something new almost every day!
I've been testing the GPU load with different settings for the CPU core usage.
I found that on a i7 with an AMD 6850M GPU, i need to leave 1 of 8 "virtual" cores free, then 2 gpu BRP processes give 80-90% gpu load. boinc suspends one cpu process if it runs 1 brp / 1 grp on the gpu, so that seems fine.
this will, however, not work on a phenom 2 x4 with a AMD 5770 gpu, here i have to free another cpu by hand for the combination BRP/GRP to run properly, otherwise i only get 30% gpu load.
if boinc would calculate the cpu usage by gpu processes differently an free one core for one or two BRP jobs (which, on my box, claim to use 0.5 cpus each while actually creating about 5% load on that cpu) and 2 cores for BRP + GRP or GRP + GRP, things should work as planned, right?
i measured power consumption of the systems, and having parts of the cpu idling around uses nearly as much power as having them do work...
if boinc would calculate the cpu usage by gpu processes differently
Ah, but it isn't BOINC that calculates how much CPU a GPU application takes. It's the scheduler (server-sided) that applies this number, and this number may depend on the relative speeds of the host's CPU and GPU.
The number is an estimate of how many CPUs the application will use. This estimate may be wildly wrong in either direction.
And even then you wouldn't want BOINC to automatically suspend the use of the X amount of CPUs/cores, as one application uses the CPU heavily and really needs the free CPU, while another on a different project uses the CPU very little and can easily piggy-back it onto the normal load of the CPU/cores.
i DID expect BRP to squeeze in between the other jobs, but that won't work.
i have to set the max number up cpus to 90%, freeing one core, then everything works fine - apart from the fact that i now have 20% idle time since BRP is not actually generating cpu load (~2%)
RE: The recommendation is
)
I agree to that. Especially in the case of laptops! In my case I did not only free one core, but disabled CPU work in general. Otherwise I would have serious temperature issues. Running my NVIDIA 610M and HD4000 is keeping it warm enough already.
RE: The only way to
)
i´d suggest that as the First Law of Crunching?
There are a lot of hosts to compare against (especially your wing men), and the log file job_log_einstein.phys.uwm.edu.txt contain a history of jobs over time, so you can see trends / changes.
Errors and invalids should be zero on an efficient system.
RE: Where do you get these
)
Rather then get into a long drawn out explanation of this or that..let me just say 'personal experience' and let it go at that, okay? As a slight testament to back that up I will let my signature come thru this time:
RE: RE: Where do you get
)
Even that doesn't stop you learning something new each day. On this project, Gary knows what he's talking about.
I realize that the GPU has to
)
I realize that the GPU has to get fed with data through the CPU, I just did not expect that a process using about 2% cpu power to not work correctly if it doesn't have a core of its own.... now, I'm wondering if it's more power efficient to run GPU tasks and have 2 of 8 cores more or less idle or run on all 8 cores and do no GPU work.
RE: RE: RE: Where do
)
And from some folks, including Gary, I do learn something new almost every day!
I've been testing the GPU
)
I've been testing the GPU load with different settings for the CPU core usage.
I found that on a i7 with an AMD 6850M GPU, i need to leave 1 of 8 "virtual" cores free, then 2 gpu BRP processes give 80-90% gpu load. boinc suspends one cpu process if it runs 1 brp / 1 grp on the gpu, so that seems fine.
this will, however, not work on a phenom 2 x4 with a AMD 5770 gpu, here i have to free another cpu by hand for the combination BRP/GRP to run properly, otherwise i only get 30% gpu load.
if boinc would calculate the cpu usage by gpu processes differently an free one core for one or two BRP jobs (which, on my box, claim to use 0.5 cpus each while actually creating about 5% load on that cpu) and 2 cores for BRP + GRP or GRP + GRP, things should work as planned, right?
i measured power consumption of the systems, and having parts of the cpu idling around uses nearly as much power as having them do work...
RE: if boinc would
)
Ah, but it isn't BOINC that calculates how much CPU a GPU application takes. It's the scheduler (server-sided) that applies this number, and this number may depend on the relative speeds of the host's CPU and GPU.
The number is an estimate of how many CPUs the application will use. This estimate may be wildly wrong in either direction.
And even then you wouldn't want BOINC to automatically suspend the use of the X amount of CPUs/cores, as one application uses the CPU heavily and really needs the free CPU, while another on a different project uses the CPU very little and can easily piggy-back it onto the normal load of the CPU/cores.
i DID expect BRP to squeeze
)
i DID expect BRP to squeeze in between the other jobs, but that won't work.
i have to set the max number up cpus to 90%, freeing one core, then everything works fine - apart from the fact that i now have 20% idle time since BRP is not actually generating cpu load (~2%)