[Mauiusers] Bug in the scheduler

Valerie Favier Valerie.Favier at bull.net
Thu Nov 16 03:30:23 MST 2006


Yesterday, I sent a mail, which explains a problem about several nodes 
and several properties. I looked the MAUI code, and I think I found a bug.
For executing a job, the selection of the node(s) is done in the 
function MReqCheckRessourceMatch (MJob.c). This function has the 
following input parameters : the job , a node and the parameter RQ 
(which contains the requirement of the job).

This selection is executed twice:
The first execution is right : the parameters: requirement, job and node 
are correct. So, the node which doesn't match the requirement (bad 
feature) is rejected.
The second execution is wrong : the parameter requirement is wrong. So, 
the node which doesn't match the requirement (bad feature) is NOT rejected.

The stack is :
MJobSelectMNL (MSched.c)
     MReqGetFNL (MJob.c)
       MReqCheckRessourceMatch (MJob.c)

The parameter requirement (RQ) is constructed in the function 
MJobSelectMNL (MSched.c).

So, can somebody help me to resolve this bug ?

My test to reproduce the problem:
I use maui-3.2.6p16 and torque 2.1.6.

I have two queues : QA and QB. The queue QA has the property A , the 
queue QB has the property B. To set the property of the queues, I use " 
resources_default.neednodes". One node NA has the property A. Two nodes 
NB1 and NB2 have the property B. ALL the nodes have 2 cpus.

I launch the following test:
qsub -l nodes=1:ppn=2+1:ppn=1  -q QB
The job runs with 2 cpus on NB2 (OK) but it runs with 1 cpu on NA. The 
property of the queue is no  respected.

Without the option -l in the qsub command, all works fine. The 
properties of the queues are always respected.

In maui.cfg , I have the following parameters :

Thanks in advance for all your help,

More information about the mauiusers mailing list