Re: [Xen-devel] Questions about Domain switching overhead for Xen 3.0

[Ryan Harper <ryanh@xxxxxxxxxx>]

* Amitayu Das <amitayudas@xxxxxxxxx> [2006-05-03 18:28]:
> Hi,
> 
> Can I ask you few quick questions about sedf scheduler available with Xen
> 3.0?
> 
> 1. Say, there are two domains apart from Dom0, each with default setting for
> slice/period length (which is (0,100) with extra flag on). Dom0 is running
> with default setting for slice/period length of (15,20). Now, is it the fact
> that irrespective of the loads experienced due to applications running
> inside the domains, a domain switching happens every 0.5 ms? We measured
> (hopefully correctly) that the context-switch (or domain-switch, to be
> precise) overhead is around 10 microsec. If the answer is yes, then would
> this mean that Xen spends 1% of its time every 0.5 ms just for switching
> domains? Is that a tolerable limit if efficiency is considered? Please let
> me know.

I don't know for sure.  Stephan may be able to help as he wrote the
SEDF scheduler for Xen.

> 
> 2. Why is the default setting for Dom0 is (15, 20) and why the default
> period length (coded in sched_sedf.c, if I'm not wrong) is 100 ms? Are there
> any justifications for these values which has not been documented?

As I understand it, the idea is to give dom0 a decent chunk of time
since it does IO on behalf of the guests.  In the remaining time dom0 is
not running, the guests will run  (extra time).  This is usually OK as
long as dom0 isn't CPU bound. 

-- 
Ryan Harper
Software Engineer; Linux Technology Center
IBM Corp., Austin, Tx
(512) 838-9253   T/L: 678-9253
ryanh@xxxxxxxxxx

_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxxxxxxxx
http://lists.xensource.com/xen-devel