Dave Hansen wrote:
You might want to try Martin Bligh's diffprofile utility. It's a bit
hard to compare 2 500-line profiles without it.
ftp://ftp.kernel.org/pub/linux/kernel/people/mbligh/tools/
Also, there have evidently been a few I/O scheduler fixes since -test3.
Please retry with -test5.
Nick Piggen wrote:
More important, are they regressions vs. previous kernels with AS?
I've checked 2.5.69mm9, 2.5.75, 2.6.0test3, and 2.6.0test5 and they all
show a
significant degradation in IO performance (>40%) when using the as
scheduler
compared to the deadline scheduler.
Configuration: IBM x440 8way with hypertheading enabled, 16GB RAM,
4 QLA2310 FC adapters attached to 2 FastT900 controllers
(112 disks total). The 112 physical disks are striped as 8 raid 0
logical disks. There are a total of 40 raw devices setup (5 per raid0
disk).
I've reattached excerpts from the database test for vmstat and
the readprofile diffs created from Martin's diffprofile utility
(diffprofile readp.as readp.dl).