I've left the task priority at 6 (below normal)...the default. It's just going to revert back to that when the next task starts anyway, and I'm not going to constantly update it.
Looking at GPU load in gpu-z, I'm finding it is highly variable. It seems to go from periods of 10-20% load up to 50 and near 60%. These periods can last 3-5 minutes each. System response is sluggish too, although that could be the gpu impacting the system rather than any cpu interference.
Indeed. Once the task is finished, the science app exits. For the next task, a new instance of the app is started by BOINC, so changing the prio will help only for the currently active app.
Still, these experiments with process prio were excellent to pinpoint the cause of the performance problem. It's just not a permanent cure.
Is there anybody with a higher gpu utilization than 60%? When running my E8400 with a FSB of 400MHz and multi of 9 I get 60% gpu usage. With a FSB of 500 and a multi of 8 I get now 66% on my gtx460.
Does somebody with a more powerful CPU experience the same (i7 without HT)?
You indeed need a powerful core to feed up the GPU. Looks like even E7200 core can not do this at full power. But may be it is because of bottleneck between RAM and GPU VRAM. So increasing then FSB helps alittle, but may be increasing the PCIE bus speed from standard 100MHz can get more affordable results.
Is there anybody with a higher gpu utilization than 60%? When running my E8400 with a FSB of 400MHz and multi of 9 I get 60% gpu usage. With a FSB of 500 and a multi of 8 I get now 66% on my gtx460.
Does somebody with a more powerful CPU experience the same (i7 without HT)?
My 8800GT peaks at about 70%. However, my 580 does not go much above 45%. I have my i7 set at 3.8 GHz.
My prefered solution for CUDA-task was to disable one of my cores. So only one CPU task is running. The second core runs the cpu-part of the Cuda-task betwenn 15 and 25% CPu-time which is 30 to 50% for one core. I didn´t wanted to set every new CUDA-task to high priority. One core would have a RAC of 1000 if I would run it 24h. With 50% used by the CUDA-task this drops to 500 and because it run it only 6h a day it would only mean that I lose 125 credits. But the CUDA-task is 8 times faster on the GPU so I "sacrifice" those 125 credits to get 1000 and don´t have to check every hour if the app is running in realtime.
If I understand you correctly, what you are saying is, in my case, I have a 6 core processor, only run 5 on CPU tasks, leaving the 'spare' one to handle the GPU tasks and that will boost my throughput to the GPU?
Im willing to try anything here to get them to ramp up to a decent workload w/o my intervention and not sluff off every task completion.
Aaron
If god meant for us not to BOINC he'd have made our #$%^%^ shorter!!
If I understand you correctly, what you are saying is, in my case, I have a 6 core processor, only run 5 on CPU tasks, leaving the 'spare' one to handle the GPU tasks and that will boost my throughput to the GPU?
Im willing to try anything here to get them to ramp up to a decent workload w/o my intervention and not sluff off every task completion.
Aaron
Yes, I set the option "on multiprocessor systems, use at most 50% of the processors". with this setting a Gw-task runs on one core and the other will be used by the cuda-task. So in your case you set the value to 83% and see what happens.
I've left the task priority
)
I've left the task priority at 6 (below normal)...the default. It's just going to revert back to that when the next task starts anyway, and I'm not going to constantly update it.
Looking at GPU load in gpu-z, I'm finding it is highly variable. It seems to go from periods of 10-20% load up to 50 and near 60%. These periods can last 3-5 minutes each. System response is sluggish too, although that could be the gpu impacting the system rather than any cpu interference.
Seti Classic Final Total: 11446 WU.
Indeed. Once the task is
)
Indeed. Once the task is finished, the science app exits. For the next task, a new instance of the app is started by BOINC, so changing the prio will help only for the currently active app.
Still, these experiments with process prio were excellent to pinpoint the cause of the performance problem. It's just not a permanent cure.
CU
HB
Is there anybody with a
)
Is there anybody with a higher gpu utilization than 60%? When running my E8400 with a FSB of 400MHz and multi of 9 I get 60% gpu usage. With a FSB of 500 and a multi of 8 I get now 66% on my gtx460.
Does somebody with a more powerful CPU experience the same (i7 without HT)?
RE: Is there anybody with a
)
As i mentioned above i get an utilization of 70% with my 8800 GTS 512.
CPU is a E7200 (no HT)
Hi! How does it compare to
)
Hi!
How does it compare to other BOINC CUDA projects?
CU
HBE
You indeed need a powerful
)
You indeed need a powerful core to feed up the GPU. Looks like even E7200 core can not do this at full power. But may be it is because of bottleneck between RAM and GPU VRAM. So increasing then FSB helps alittle, but may be increasing the PCIE bus speed from standard 100MHz can get more affordable results.
RE: Is there anybody with a
)
My 8800GT peaks at about 70%. However, my 580 does not go much above 45%. I have my i7 set at 3.8 GHz.
My prefered solution for
)
My prefered solution for CUDA-task was to disable one of my cores. So only one CPU task is running. The second core runs the cpu-part of the Cuda-task betwenn 15 and 25% CPu-time which is 30 to 50% for one core. I didn´t wanted to set every new CUDA-task to high priority. One core would have a RAC of 1000 if I would run it 24h. With 50% used by the CUDA-task this drops to 500 and because it run it only 6h a day it would only mean that I lose 125 credits. But the CUDA-task is 8 times faster on the GPU so I "sacrifice" those 125 credits to get 1000 and don´t have to check every hour if the app is running in realtime.
If I understand you
)
If I understand you correctly, what you are saying is, in my case, I have a 6 core processor, only run 5 on CPU tasks, leaving the 'spare' one to handle the GPU tasks and that will boost my throughput to the GPU?
Im willing to try anything here to get them to ramp up to a decent workload w/o my intervention and not sluff off every task completion.
Aaron
If god meant for us not to BOINC he'd have made our #$%^%^ shorter!!
RE: If I understand you
)
Yes, I set the option "on multiprocessor systems, use at most 50% of the processors". with this setting a Gw-task runs on one core and the other will be used by the cuda-task. So in your case you set the value to 83% and see what happens.