[concurrency-interest] padding in Exchanger
dl at cs.oswego.edu
Tue Jan 17 10:52:10 EST 2012
On 01/17/12 10:45, Michael Barker wrote:
> I was about to ask the same question, my understanding of false
> sharing is that it is triggered by writes to the same cache line by
> multiple cores resulting in excessive RFO traffic. I.e. ownership of
> the contented cache line moves frequently between cores. If caches
> lines are fetched together but written back independently would it
> still be a problem?
Empirically, yes. If you are curious about your own machines,
one way to check this out is to vary the "ASHIFT" constant in the
new version of Exchanger (which you'd then need to compile/build).
The default is 7 (=> (1 << 7) == 128) and then run some of
our perf tests like ExchangeLoops.
More information about the Concurrency-interest