Ryan Cox
2014-09-26 12:27:30 UTC
It was great to see so many of you at Slurm User Group Meeting. We
received several questions after our presentation and wanted to clarify
some things.
One of our "possible concerns" was about a "tiny user in a very active
account". To clarify, this is the scenario we were mentioning. We have
a very active account (let's call them "acctbob") that is the account
with the highest usage of anyone on our systems. Occasionally acctbob
has a user (userjane) who only rarely runs and wants to get some work
done. userjane has trouble running because she is in acctbob, the
account with the lowest fairshare of anyone. Some may see this as a
problem and wish to do something about it, to which we suggest an
optional QOS with higher priority and a high usage factor that can be
used when she wants to run quickly but is okay being "penalized" for
that usage. Alternatively userjane may be given access to a secondary
QOS with tight limits. Or, you could do the same as us and say "sorry,
that's the way it is". This problem mostly exists no matter which
algorithm you choose.
Decay factor works in our algorithm. Decay factor is a mostly
orthogonal concept that fairshare algorithms happen to include.
One other thing to note for people who weren't at the meeting is that
Fair Tree replaces Level-Based, the other algorithm we posted about a
few months ago.
Please let us know if you have any other questions.
Ryan and Levi
P.S. Several of you asked us to send the slides to you even before Moe
posts them next week. In case I missed anyone, here is a link:
https://drive.google.com/file/d/0B8dVHMccGpAJcTZCdmE3Wjhya3M/edit?usp=sharing
received several questions after our presentation and wanted to clarify
some things.
One of our "possible concerns" was about a "tiny user in a very active
account". To clarify, this is the scenario we were mentioning. We have
a very active account (let's call them "acctbob") that is the account
with the highest usage of anyone on our systems. Occasionally acctbob
has a user (userjane) who only rarely runs and wants to get some work
done. userjane has trouble running because she is in acctbob, the
account with the lowest fairshare of anyone. Some may see this as a
problem and wish to do something about it, to which we suggest an
optional QOS with higher priority and a high usage factor that can be
used when she wants to run quickly but is okay being "penalized" for
that usage. Alternatively userjane may be given access to a secondary
QOS with tight limits. Or, you could do the same as us and say "sorry,
that's the way it is". This problem mostly exists no matter which
algorithm you choose.
Decay factor works in our algorithm. Decay factor is a mostly
orthogonal concept that fairshare algorithms happen to include.
One other thing to note for people who weren't at the meeting is that
Fair Tree replaces Level-Based, the other algorithm we posted about a
few months ago.
Please let us know if you have any other questions.
Ryan and Levi
P.S. Several of you asked us to send the slides to you even before Moe
posts them next week. In case I missed anyone, here is a link:
https://drive.google.com/file/d/0B8dVHMccGpAJcTZCdmE3Wjhya3M/edit?usp=sharing