Re: [boinc_dev] #GPUs #cores
Bernd: Please verify this experimentally; I'm not convinced. -- David On 15-Apr-2011 1:12 AM, Bernd Machenschalk wrote: On 15.04.11 08:00, David Anderson wrote: The client tries to use all GPUs, even if this overcommits the CPUs. The assumption is that this maximizes throughput. Is there evidence that it doesn't? There are two types of GPU applications: such that run almost entirely on the GPU and just need a small fraction of a CPU core, and such that use the GPU mainly as a coprocessor for certain parts of the computation (e.g. FFT) and require a full CPU core. I'd say the assumption is true for the former, but not for the latter. So similar to the distinction made between these two types of Apps in assigning process priority, I'd do this in resource counting, too. Best, Bernd ___ boinc_dev mailing list boinc_dev@ssl.berkeley.edu http://lists.ssl.berkeley.edu/mailman/listinfo/boinc_dev To unsubscribe, visit the above URL and (near bottom of page) enter your email address. ___ boinc_dev mailing list boinc_dev@ssl.berkeley.edu http://lists.ssl.berkeley.edu/mailman/listinfo/boinc_dev To unsubscribe, visit the above URL and (near bottom of page) enter your email address.
Re: [boinc_dev] #GPUs #cores
The client tries to use all GPUs, even if this overcommits the CPUs. The assumption is that this maximizes throughput. Is there evidence that it doesn't? -- David On 12-Apr-2011 4:51 AM, Bernd Machenschalk wrote: Hi! We're experimenting with running BOINC on a cluster of GPU nodes. Our application takes a full core per NVidia GPU (avg_ncpus = 1.0). The BOINC Client is told to use only one CPU core (for now), i.e.ncpus1/ncpus in cc_config.xml. However the Client starts as many tasks as there are GPUs on that node. When scheduling GPU tasks, does the Client ignore the number of available cores, expecting that there will always be more cores than GPUs? If so I'd consider this a bug. Best, Bernd ___ boinc_dev mailing list boinc_dev@ssl.berkeley.edu http://lists.ssl.berkeley.edu/mailman/listinfo/boinc_dev To unsubscribe, visit the above URL and (near bottom of page) enter your email address. ___ boinc_dev mailing list boinc_dev@ssl.berkeley.edu http://lists.ssl.berkeley.edu/mailman/listinfo/boinc_dev To unsubscribe, visit the above URL and (near bottom of page) enter your email address.
[boinc_dev] #GPUs #cores
Hi! We're experimenting with running BOINC on a cluster of GPU nodes. Our application takes a full core per NVidia GPU (avg_ncpus = 1.0). The BOINC Client is told to use only one CPU core (for now), i.e. ncpus1/ncpus in cc_config.xml. However the Client starts as many tasks as there are GPUs on that node. When scheduling GPU tasks, does the Client ignore the number of available cores, expecting that there will always be more cores than GPUs? If so I'd consider this a bug. Best, Bernd ___ boinc_dev mailing list boinc_dev@ssl.berkeley.edu http://lists.ssl.berkeley.edu/mailman/listinfo/boinc_dev To unsubscribe, visit the above URL and (near bottom of page) enter your email address.