Re: Context switch time regression...

From: David Mosberger <davidm_at_napali.hpl.hp.com>
Date: 2005-03-15 05:03:38
>>>>> On Mon, 14 Mar 2005 15:16:46 +1100, Peter Chubb <peterc@gelato.unsw.edu.au> said:

  Peter> According to LMbench2 on a 2-way 900MHz Itanium 2:

  Peter> Context switching 0K - times in microseconds - smaller is better
  Peter> ----------------------------------------------------------------
  Peter> OS                               2p/0k      4p/0k      8p/0k
  Peter> ctx swtch  ctx swtch  ctx swtch
  Peter> -----------------------------  ---------  ---------  ---------
  Peter> Linux 2.5.72                   1.5700000  1.6200000  1.7700000
  Peter> Linux 2.6.0                      15.0600  8.6900000  8.8600000
  Peter> Linux 2.6.11                     17.1900    17.2100    10.5400

  Peter> I have to go now, but I'll try to track this down tomorrow.

I have noticed the same a while ago.  Someone from Intel (Ken, IIRC)
investigated and confirmed that it is due to a scheduler change.
Basically, the scheduler is now much more aggressive about
loading-balancing to an idle CPU.  Try running LMbench as "taskset 1
make rerun" and you'll see the effect.  This has resulted in a
discussion with Carl Staelin (co-author and current maintainer of
LMbench) as to how LMbench could control for such issues.  It's not
finished yet, but hopefully a future version of LMbench3 will
(optionally) allow you to control the affinity of processes so you can
factor out scheduling effects if desired.

	--david
-
To unsubscribe from this list: send the line "unsubscribe linux-ia64" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Received on Mon Mar 14 13:14:06 2005

This archive was generated by hypermail 2.1.8 : 2005-08-02 09:20:36 EST