Many projects do have feature-specific application versions, and the plan_class
mechanism supports it already.
Unfortunately, the rule under RuntimeEstimation seems to prioritise "forbid
sending incompatible version" (which is important, of course). The scheduler
tests the full range of compatible apps, like SSE2/SSE3/SSSE3/SSE4.1/AVX/AVX2
for CPUs, CUDA 3/4/5/6/7/8/9 for GPUs, and records the speed of each.
Sometimes, some other variable transient in task runtime causes the scheduler
to latch onto a less-optimised version thinking it is fastest, and it is
difficult to break out of that misapprehension - which slightly reduces the
effectiveness of the whole concept.
On Tuesday, 16 January 2018, 9:52, Vitalii Koshura
<[email protected]> wrote:
Hello @all,
As far as I know there is no possibility for boinc client to ask project
server for the most optimized application based on available set of
instructions (e.g. AVX), am I right?
This feature can decrease the time necessary to make calculations of the
work unit on new CPUs because of using more optimized code.
>From another hand this will not break backward compatibility for older CPUs
that have no such extensions and can run less optimized code only.
Also this feature can be extended to GPUs (probably).
>From my POV this is rather good feature. What do you think, guys?
Best regards,
Vitalii Koshura
_______________________________________________
boinc_dev mailing list
[email protected]
https://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
[email protected]
https://lists.ssl.berkeley.edu/mailman/listinfo/boinc_dev
To unsubscribe, visit the above URL and
(near bottom of page) enter your email address.
