[Mauiusers] fairshare - a better way to calculate percentages?

Michael Barnes barnes at jlab.org
Wed Jun 17 12:07:05 MDT 2009


On Tue, Jun 16, 2009 at 07:28:03AM +0200, Bas van der Vlies wrote:
> I knew there is a patch for it, see:
>   * http://www.mail-archive.com/mauiusers@supercluster.org/msg00321.html

My manager has asked me to look into the Maui fairshare because when all
of the active users are over their fairshare their fairshare weighting
becomes equal.  He said it quite well:

On Fri, Mar 27, 2009 at 10:26:36AM -0400, Chip Watson wrote:
> Michael,
>
> Maui has the following known bug:
>
> If all projects that have pending jobs are over their fairshare, and
> the other projects are not being used, then Maui takes the unused
> fairshare and divides it EQUALLY among all projects. It does NOT
> divide it up in proportion to the other projects fairshare. E.g.
> Project A share is 80, B is 19, C is 1. A goes inactive, and so the 80
> is split equally between B and C. So B ends up with 59, and C with 41,
> changing the ratio between B and C from 19:1 to 3:2.
>
> What I'm wondering now is if the job soft limit is also subverted by
> this non-fairshare of unused fairshare. In other words, soft limit is
> only imposed if there is no one ready to run who is below their quota.
>
> Can you dig into the source code to figure this out?

In looking at the code in MPriority.c, MFS.c, and MQueue.c, I found
that the fairshare code was a) being clamped to 0 if negative and
b) the tools like diagnose -p do not show the actual values of the
priority (they report negative numbers when they are never used). Also,
the priorities go from doubles, to longs, to ints, which could cause
problems as well.

We run a slightly older version of Maui, and if these issues has been
fixed, I would immediately upgrade.


Is anyone else aware of these issues?

> Just a question to the cluster resources people is this the right  
> approach and if yes will the patch be applied if it is a config option?

I was wondering the same thing.  I was thinking of adding a FSPAD config
value to skew all of the FS values up so that they will not ever go
negative.  I have not thought too much about it, but it might be an easy
fix.

-mb

-- 
+-----------------------------------------------
| Michael Barnes
|
| Thomas Jefferson National Accelerator Facility
| 12000 Jefferson Ave.
| Newport News, VA 23606
| (757) 269-7634
+-----------------------------------------------


More information about the mauiusers mailing list