Ronald Moesbergen, on 08/03/2009 01:15 PM wrote:
2009/7/31 Vladislav Bolkhovitin <vst@xxxxxxxx>:
OK, as I expected, on the SCST level everything is clear and the forced
ordering change didn't change anything.
But still, a single read stream must be the fastest from single thread.
Otherwise, there's something wrong somewhere in the I/O path: block layer,
RA, I/O scheduler. And, apparently, this is what we have and should find out
the cause.
Can you check if noop on the target and/or initiator makes any difference?
Case 5 with 1 and 2 threads will be sufficient.
That doesn't seem to help:
client kernel: 2.6.26-15lenny3 (debian)
server kernel: 2.6.29.5 with readahead-context, blk_run_backing_dev
and io_context, forced_order
With one IO thread:
5) client: default, server: default (server noop, client noop)
blocksize R R R R(avg, R(std R
(bytes) (s) (s) (s) MB/s) ,MB/s) (IOPS)
67108864 17.612 21.113 21.355 51.532 4.680 0.805
33554432 18.329 18.523 19.049 54.969 0.891 1.718
16777216 18.497 18.219 17.042 57.217 2.059 3.576
With two threads:
5) client: default, server: default (server noop, client noop)
blocksize R R R R(avg, R(std R
(bytes) (s) (s) (s) MB/s) ,MB/s) (IOPS)
67108864 17.436 18.376 20.493 54.807 3.634 0.856
33554432 17.466 16.980 18.261 58.337 1.740 1.823
16777216 18.222 17.567 18.077 57.045 0.901 3.565
And with client cfq, server noop?
Ronald.
--
To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
