* Nick Piggin <nickpiggin@xxxxxxxxxxxx> wrote:There's one important bit missing from that graph, the 2.6.23-SCHED_BATCH values. Without that we can't tell how much improvement is from sched-devel and how much from SCHED_BATCH. Clearly 2.6.23 is better than 2.6.22.any in this test, the locking issues seem to dominate that difference to the point that nothing else would be informative.
;) I think you snipped the important bit:
"the peak is terrible but it has virtually no dropoff and performs better under load than the default 2.6.21 scheduler." (verbatim)
hm, i understood that peak remark to be in reference to FreeBSD's scheduler (which the FreeBSD guys are primarily interested in obviously), not v2.6.21 - but i could be wrong.
In any case, there is indeed a regression with sysbench and a low number of threads, and it's being fixed. The peak got improved visibly in sched-devel:
http://people.redhat.com/mingo/misc/sysbench-sched-devel.jpg
but there is still some peak regression left, i'm testing a patch for that.