P.S. While I have your attention, let me provide some related info. for i in 1 2 3 4 5; do ipcscale/sem-waitzero -t 30 -c 1,8,16,24,32,40,48,56,64 -i 1 -m 0 -d 2 -f 2>&1|tee -a log-blah; done 3.0-rt (SLE11-SP3-RT with sem backports) scales perfectly to 64 cores on my 8 node (alas, 1 memory node) DL980 rt test box, whether configured PREEMPT_RT/1000 or NOPREEMPT/250. (livelock bug is irrelevant) 3.10.10-rt7 scales horridly, is erratic as all hell whether configured PREEMPT_FULL/1000 or NOPREEMPT/250. Ditto pure 3.10.10 NOPREEMPT/250. (3.10-rt kernel is where I started, Manfred asked me to try something, I happened to have this kernel on box to take a peek at future workhorse) Master NOPREEMPT/250 as of yesterday with same config taken from tests above scales perfectly. So, that _alleges_ that 3.10.10, a.k.a next longterm stable kernel, is sick, missing something. It appears something went wonky post 3.0, but that's only possibly interesting for finding the origin for -stable iff the thing turns out to be _really_ stable/bisectable in the first place, which I'm going to find out, hopefully today. It _looks_ stable. Would be nice if someone would confirm/deny bad big box behavior, but... -Mike -- To unsubscribe from this list: send the line "unsubscribe linux-rt-users" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
