Thanks.
This was in relation to some odd behaviour of the CreditNew system where
GPU tasks would get much higher variance than CPU tasks for the same
workload. Systems where the GPU was shared with one or more other
processes or where the GPU was not fully utilized because another CPU
process was starving the GPU thread on the CPU generally claimed more
credit. The use of wall-time for tracking explains the results seen in
the test.
Just a side note:
I may be wrong but I believe "GPU time" or "context switch accounting"
or just "accounting" was added in CUDA 5.5 and allows you to get the
accumulative number of millisecs a process's context was loaded in the
active area of the GPU over the lifetime of that process along with the
average utilization coefficient for that process during that time. This
is probably the closest we will get to any definition of "GPU time".
This data may be available through the NVML API (and through that the
'nvidia-smi' command). However, the format and access requirements to
that data was focused heavily on the need for GPU Cloud computing
cluster owners to bill their users and as such may not be suitable for
BOINC at all.
Maybe we need to gently prod Nvidia (and friends) to allow us to enable
and query this info from within a process itself at some point in the
future? - the hardware counters required to do so seem to be available
on the cards nowadays. It could be a good way to increase accuracy when
tracking the use of GPU resources for BOINC going forward.
-- Janus
On 01-06-2015 20:23, David Anderson write:
It's still the case.
The notion of "GPU time" is not clear,
since apps in general use a fraction of the GPU's cores.
-- David
On 01-Jun-2015 10:57 AM, Janus wrote:
On 03-09-2009 kl. 18:25 David Anderson wrote:
Nicolás Alvarez wrote:
No, the client doesn't measure "GPU time" at all. Is it even
possible to
get that information from nvidia APIs?
I couldn't find such an API.
[SNIP]
Sorry for necroing this old thread from 2009 but is this still the case?
-- Janus
_______________________________________________
boinc_dev mailing list
[email protected]
http://lists.ssl.berkeley.edu/mailman/listinfo/boinc_dev
To unsubscribe, visit the above URL and
(near bottom of page) enter your email address.
