[torquedev] PPN/Node state bug?
SimonT at mail.muni.cz
Fri Sep 11 11:21:27 MDT 2009
> On Fri, Sep 11, 2009 at 05:04:41PM +0200, Simon Toth alleged:
>> I just run into weird behavior. The node state seems to be updated only
>> if all of the nodes processors are full (then it switches to
>> job-exclusive or job-sharing).
> Yes, that's how it has always worked. When all CPUs are allocated, then the
> node state gets the job-exclusive bit.
So when I request an exclusive job, its actually not exclusive?
>> Is this normal? How can the scheduler determine if the node can run more
>> jobs if the server does not report if the node is running exclusive or
>> sharing jobs?
> How would the server know if the jobs are exclusive or shared? That's for the
> scheduler to know.
Uh? The standard scheduler doesn't determine this at all, its the
servers job. He is parsing the nodespec and assigning nodes to the job.
In my scheduler I keep this functionality on (I just add the target
hostname to each part of the nodespec) because I need the server to
verify the nodespec anyway. The state of nodes might have changed from
time I downloaded the information from the server, so the server needs
to check my assumptions.
Btw. how would the scheduler know if the server doesn't report this
information (he standard server doesn't even report free cpus)? All the
scheduler has is information from the server.
Mgr. Simon Toth
160 00 Praha 6
-------------- next part --------------
A non-text attachment was scrubbed...
Size: 5207 bytes
Desc: S/MIME Cryptographic Signature
Url : http://www.supercluster.org/pipermail/torquedev/attachments/20090911/772409e5/attachment.bin
More information about the torquedev