On Thu, Jul 12, 2012 at 03:40:14AM -0500, Mark Nipper wrote: > On 12 Jul 2012, Brian Candler wrote: > > And I forgot to add: since a KVM VM is a userland process anyway, I'd expect > > a big performance gain when KVM gets the ability to talk to libglusterfs to > > send its disk I/O directly, without going through a kernel mount (and hence > > bypassing the kernel cache). It looks like this is being developed now: > > http://lists.gnu.org/archive/html/qemu-devel/2012-06/msg01745.html > > You can see the performance figures at the bottom of that post. > > Something concerns me about those performance figures. > If I'm reading them correctly, the normal fuse mount performance > is about what I was seeing, 2-3MB. And now bypassing everything, > libglusterfs is still capping out a little under 20MB/s. I read it as: (aggrb) - base 72.9MB/s - fuse bypass (libglusterfs) 66.MB/s (minb) - base 18.2MB/s - fuse bypass 16.6MB/s (maxb) - base 18.9MB/s - fuse bypass 17.8MB/s etc.
