Every year someone shows up trying to convince everyone that there
will be a giant speed up if BOINC implements CPU affinity, based on
theory and not practice. Please try setting CPU affinity manually when
WUs start, and let us know how much faster they finish.
Large enough performance improvements (measured by experiment) on real
BOINC apps is more likely to motivate change than citing papers.
Enviado desde mi iPod
El 11/03/2010, a las 14:06, "Charles Elliott" <[email protected]>
escribió:
If anyone is interested there is a good explanation for why BOINC
applications experience a speed-up
when each application is assigned an affinity for one CPU (and by
implication why it does not work for
single-CPU hyperthreading) here:
Fedorova, A., S. Blagodurov, et al. (2010). "Managing contention for
shared
resources on multicore
processors." Commun. ACM 53(2): 49-57.
Essentially, one can infer from the article that the speed-up is due
to
fewer cache misses and
reduced contention for the front-side bus and the memory controller.
Charles Elliott
_______________________________________________
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.
_______________________________________________
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.
