RE: Best queue system for myrinet?

"Tierney, Craig" <[email protected]>
Newsgroups gmane.network.myrinet.general
Message-ID <22699FB2708ED611ADA100105AC5C9A05B02C2@hpti_main.hpti.com>
 

-----Original Message-----
From: Thomas Davis
To: Bill Broadley
Cc: [email protected]
Sent: 11/18/02 1:13 PM
Subject: Re: [Myrinet] Best queue system for myrinet?

Bill Broadley wrote:
> I'm reinstalling a 32 node (64 cpu) myrinet cluster.
> 
> I'm looking for a batch queue system, but want to avoid the problem
> I've had in the past (although this info maybe somewhat dated).
>> 
>> Basically I want to schedule my cluster as 64 cpu's, but if I
>> tell the queue system I have more then 32 cpu it needs to know
>> that the first job on a given node gets Myrinet channel 2, and the
> second 
>> job on the same node gets Myrinet channel 4.
>> 
>> What is working well for people?
>> 
>
>We have been doing some batch queue testing with ESP2, becasue we are 
>greatly disappointed with the reliability (and ability of) OpenPBS.
>
>
>We've tested to date the following:
>
>1) OpenPBS w/Maui
>2) SGE (free edition)
>3) Platform's LSF
>4) Maui/ME
>
>The problems of each:
>
>1) Hangs; has tendency to put a job in a wierd state, that sends 
>thousands of email messages saying so; only way to clear is stop the 
>master, and delete the job file by hand.
>
It hangs because the server tries to talk to the nodes actively,
and if the node doesn't respond the server hangs.  Try a 
strace on the server process and see what is happening.

>
>2) Scheduler doesn't work very well - wants to run all the small jobs 
>first, priorities only apply after jobs are sorted by size.
>

I have not seen this with SGE.  I do not have this problem.  I have
seen where the scheduler does not schedule the oldest job though 
(all same size, all same priority).


Craig
lmpx.com only provides a reader for public news (NNTP) servers. It is not affiliated with the servers or forums shown here and is not responsible for the content of articles, which is written by their respective authors.