From mboxrd@z Thu Jan 1 00:00:00 1970 From: "John Hawkes" Date: Wed, 24 Aug 2005 16:09:40 +0000 Subject: Re: [PATCH 2.6.13-rc6] cpu_exclusive sched domains build fix Message-Id: <006a01c5a8c6$3c2006d0$6600a8c0@PCJohn> List-Id: References: <20050824111510.11478.49764.sendpatchset@jackhammer.engr.sgi.com> <20050824112640.GB5197@in.ibm.com> MIME-Version: 1.0 Content-Type: text/plain; charset="us-ascii" Content-Transfer-Encoding: 7bit To: dino@in.ibm.com, Paul Jackson Cc: paulus@samba.org, Andrew Morton , nickpiggin@yahoo.com.au, linux-ia64@vger.kernel.org, linux-kernel@vger.kernel.org, torvalds@osdl.org, mingo@elte.hu From: "Dinakar Guniguntala" > Can we hold on to this patch for a while, as I reported yesterday, > this hangs up my ppc64 box on doing rmdir on a exclusive cpuset. > Still debugging the problem, hope to have a fix soon, Thanks Paul's patch simply constrains the scope of cpuset configurations that will invoke the "dynamic sched domains" functionality, which means that some cpu-exclusive (a.k.a. "isolated") cpusets will continue to have the 2.6.12-and-earlier behavior of being periodically examined by the CPU Scheduler in load-balancing activities. That is, Paul's patch simply reverts cpuset/sched domain behavior to pre-2.6.13 status (for some cpusets). The pre-2.6.13 non-"dynamic sched domains" behavior will in fact produce bad load-balancing behavior if a cpu-exclusive cpuset is so heavily loaded with executing processes, all pinned to the cpu(s) in the cpuset, that the other cpus in the system see this cpu(s)/node as the most-heavily-loaded and just focus on it during load-balancing -- which would be futile, of course, since the processes pinned to this highest-load cpu (and node) cannot be offloaded. Since load-balancing looks only at the most-heavily-loaded cpu as a cpu to offload, this means that all system load-balancing would be effectively turned off. John Hawkes