Re: Schedulers benchmark - Was: [ANNOUNCE][RFC] PlugSched-5.2.4 for2.6.12 and 2.6.13-rc6

[Peter Williams]

Lee Revell wrote:
> On Fri, 2005-08-19 at 05:09 +0200, Michal Piotrowski wrote:
>
> > Hi,
> > here are interbench v0.29 resoults:
>
>
> The X test under simulated "Compile" load looks most interesting.
>
> Most of the schedulers do quite poorly on this test - only Zaphod with
> default max_ia_bonus and max_tpt_bonus manages to deliver under 100ms
> max latency.  As expected with interactivity bonus disabled it performs
> horribly.
>
> I'd like to see some results with X reniced to -10.  Despite what the
> 2.6 release notes say, this still seems to make a difference.

Even spa_no_frills, which does absolutely nothing to help interactive 
(or other special interest) tasks, can cope in these circumstances as 
illustrated by these results from my (relatively old) SMP machine show:

--- Benchmarking simulated cpu of X nice -10 in the presence of 
simulated ---
Load	Latency +/- SD (ms)  Max Latency   % Desired CPU  % Deadlines Met
None	   0.01 +/- 0.129          2		 100	       99.3
Video	  0.007 +/- 0.0818         1		 100	       99.3
Burn	  0.006 +/- 0.0817         1		 100	       99.3
Write	  0.033 +/- 0.271          3		99.3	         98
Read	  0.046 +/- 0.337          3		98.4	         97
Compile	  0.023 +/- 0.208          2		99.3	       98.3
Memload	  0.043 +/- 0.31           3		98.1	         97

This machine isn't directly comparable with Michal's so for comparison 
here are the results from "out of the box" Zaphod on the same machine:

--- Benchmarking simulated cpu of X in the presence of simulated ---
Load	Latency +/- SD (ms)  Max Latency   % Desired CPU  % Deadlines Met
None	   0.02 +/- 0.2            2		99.3	       98.7
Video	  0.007 +/- 0.0818         1		 100	       99.3
Burn	  0.023 +/- 0.208          2		99.3	       98.3
Write	  0.147 +/- 0.949         12		94.7	       93.2
Read	  0.033 +/- 0.258          2		98.7	       97.7
Compile	   2.94 +/- 10.7         105		76.8	       71.6
Memload	  0.017 +/- 0.153          2		 100	       98.7

As you can see there's evidence in these numbers the file writes are 
implicated in the bad numbers for the Compile load (which is a mixture 
of Burn, Read, Write and (I think) Memload).  So testing with different 
I/O schedulers might be interesting.

Peter
--
Peter Williams                                   pwil3058@xxxxxxxxxxxxxx

"Learning, n. The kind of ignorance distinguishing the studious."
 -- Ambrose Bierce
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/