[Xen-devel] high I/O cost for drbd and xen?

["Tom Hibbert" <tom@xxxxxxxxx>]

Hi all
 
Just run some benchmarks on my new failover cluster and found some
alarming results. I've previously observed that xen typically introduces
almost zero IO overhead. I ran some tests with bonnie -x 10 on my drbd
mirrors and compared the averages between dom0 and dom1 to see the
virtualisation cost. I was surprised (and alarmed) to see that there was
a fairly high cost for read operations (high being >10%). I was very
concerned to see the 80% cost for get_block on the failover node.
 
Could this be a scheduler issue? drbd likes to be high priority - it's a
high availability service after all. Perhaps the hypervisor is blocking
it
or not allowing it its normal share of cpu time? Perhaps its giving CPU
time from drbd on dom0 to bonnie on dom1, causing the reported slowdown?
 
Tom

benchmarks.csv
Description: benchmarks.csv