I'm running 2.6.29.4-rt16 with CONFIG_PREEMPT_RT=y on an Intel Core 2 Quad Q9650 and I have recently noticed something odd. (I've also seen this on 2.6.29.2-rt11, and I have not tried any other -rt versions so I don't know how new this is in the series of -rt kernel versions.) On an idle machine, I open four new xterms, and then in each xterm I type: while : ; do : ; done I'm watching what happens in another window that is running htop. Sometimes (though not often) I'll see all four CPUs go to 100%, and four shell processes listed as running at 100%. More often, one or more CPUs remain essentially idle, and from the CPU% displayed in htop I'll see that some of my looping bash processes appear to be sharing a CPU, each getting around half or a third of the CPU%. In a few tests I've never seen all four on the same core with three idle CPUs, but I usually see one or two idle CPUs with the CPU% of my looping bashes revealing which have a CPU to themselves and which are sharing. It is as if the processes somehow get arbitrarily assigned to a core and never get moved even though there's contention for the core they are stuck on and one or two other idle cores. If I ^C the loops and restart them, the pattern seems to be the same. But if I exit the xterms and start up four new xterms, I'll get some new pattern, which is repeatable as many times as I like with those particular xterm/bash combinations. (I do disable CPU speed switching by running for i in 0 1 2 3 ; do sudo cpufreq-set -f 3g -c $i ; done shortly after booting, but I have no reason to think that is relevant.) Does anyone else see this same weirdness? Is this an actual bug in -rt or am I wrong to expect the -rt kernel to migrate processes if some processors are busy and others idle? (I know I could try distributing processes on cores manually using taskset, but it surprises me I should have to do this to get all CPUs working.) (I don't see this on kernels without the -rt patch.) -Tim Shepard shep@xxxxxxxxxxxx -- 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
