Re: [BENCHMARK] I/O regression after 2.6.0-test5

[John Hawkes <hawkes@babylon.engr.sgi.com>]

I ran AIM7 on a 64p Altix NUMA (Itanium 2 Madison CPUs, 1.5 GHz), with an
XFS filesystem on each of 99 fiberchannel disks, with two different flavors
of AIM7 workloads, on both 2.6.0-test5 and 2.6.0-test8.  One workload shows
that -test8 is half the throughput of -test5.  The other workload shows
that -test8 is 22% faster.

The first workload is the default AIM7 "multiuser/shared system" workload.
2.6.0-test8 shows a peak throughput of 52% of -test5 -- it drops in half.
The -test8 peak at 64p is roughly the same as at 8p and 4p.  This
bottleneck cannot be attributed solely to I/O hardware, since our SGI
ProPack kernel (based upon 2.4.21) achieves a peak throughput at 64p that
is 30% higher than 2.6.0-test5's peak at 64p.

The second workload takes the same "multiuser/shared system" workload,
but turns off the O_SYNC flag for the three sync_* subtests.  This
generally produces higher peak throughput values than when using the
default O_SYNC, and the peak throughput values scale significantly better
as CPU counts increase.  Here the -test8 kernel is 22% *faster* than
-test5 at 64p (and 8x faster for no-O_SYNC than with-O_SYNC).

John Hawkes