From mboxrd@z Thu Jan 1 00:00:00 1970 From: George Dunlap Subject: [PATCH] docs: Document scheduler-related Xen command-line options Date: Mon, 24 Sep 2012 16:19:57 +0100 Message-ID: <7e332fd064fac8d9d1ce.1348499997@elijah> Mime-Version: 1.0 Content-Type: text/plain; charset="us-ascii" Content-Transfer-Encoding: 7bit Return-path: List-Unsubscribe: , List-Post: List-Help: List-Subscribe: , Sender: xen-devel-bounces@lists.xen.org Errors-To: xen-devel-bounces@lists.xen.org To: xen-devel@lists.xensource.com Cc: george.dunlap@eu.citrix.com List-Id: xen-devel@lists.xenproject.org # HG changeset patch # User George Dunlap # Date 1348499935 -3600 # Node ID 7e332fd064fac8d9d1cea904d5236c8d74389194 # Parent 7b045d43e59dcb42340097058502bf456e151180 docs: Document scheduler-related Xen command-line options Signed-off-by: George Dunlap diff --git a/docs/misc/xen-command-line.markdown b/docs/misc/xen-command-line.markdown --- a/docs/misc/xen-command-line.markdown +++ b/docs/misc/xen-command-line.markdown @@ -732,12 +732,34 @@ Choose the default scheduler. ### sched\_credit\_tslice\_ms > `= ` +Set the timeslice of the credit1 scheduler, in milliseconds. The +default is 30ms. Reasonable values may include 10, 5, or even 1 for +very latency-sensitive workloads. + ### sched\_ratelimit\_us > `= ` +In order to limit the rate of context switching, set the minimum +amount of time that a vcpu can be scheduled for before preempting it, +in microseconds. The default is 1000us (1ms). Setting this to 0 +disables it altogether. + ### sched\_smt\_power\_savings > `= ` +Normally Xen will try to maximize performance and cache utilization by +spreading out vcpus across as many different divisions as possible +(i.e, numa nodes, sockets, cores threads, &c). This often maximizes +throughput, but also maximizes energy usage, since it reduces the +depth to which a processor can sleep. + +This option inverts the logic, so that the scheduler in effect tries +to keep the vcpus on the smallest amount of silicon possible; i.e., +first fill up sibling threads, then sibling cores, then sibling +sockets, &c. This will reduce performance somewhat, particularly on +systems with hyperthreading enabled, but should reduce power by +enabling more sockets and cores to go into deeper sleep states. + ### serial\_tx\_buffer > `= `