>>> On 18.12.15 at 17:08, <malcolm.cross...@citrix.com> wrote: > The per domain grant table read lock suffers from significant contention when > performance multi-queue block or network IO due to the parallel > grant map/unmaps/copies occurring on the DomU's grant table. > > On multi-socket systems, the contention results in the locked compare swap > operation failing frequently which results in a tight loop of retries of the > compare swap operation. As the coherency fabric can only support a specific > rate of compare swap operations for a particular data location then taking > the read lock itself becomes a bottleneck for grant operations. > > Standard rwlock performance of a single VIF VM-VM transfer with 16 queues > configured was limited to approximately 15 gbit/s on a 2 socket Haswell-EP > host. > > Percpu rwlock performance with the same configuration is approximately > 48 gbit/s. > > Oprofile was used to determine the initial overhead of the read-write locks > and to confirm the overhead was dramatically reduced by the percpu rwlocks. > > Signed-off-by: Malcolm Crossley <malcolm.cross...@citrix.com>
Reviewed-by: Jan Beulich <jbeul...@suse.com> _______________________________________________ Xen-devel mailing list Xen-devel@lists.xen.org http://lists.xen.org/xen-devel