Hi
Just a quick suggestion to change the client to use round-robin order
when starting up new GPU workunits rather than the current
fill-each-card-before-proceeding-order. An even smarter solution (but
probably more complex) would be to always add work to the least used card.
The reasoning is that when work-fetch is enabled for a system with many
GPUs and many WU instances per GPU it may take a while to request enough
work from the scheduler to fill all the way to the last device. If the
instances are instead spread across different pieces of hardware from
the get-go they will get higher performance (lower WU return latency as
seen from the project side) than the same number of instances running on
just one board.
-- Janus
_______________________________________________
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.