Re: [HTCondor-devel] STARTD_SLOT_ATTRS and new ClassAd syntax


Date: Fri, 05 Feb 2016 11:16:58 +0100
From: Joan Piles <jpiles@xxxxxxxxxxxxxxxx>
Subject: Re: [HTCondor-devel] STARTD_SLOT_ATTRS and new ClassAd syntax
Hi Greg,

I had indeed thought about machine resources, but that would only work if there is a known number of resources (e.g. network bandwith), and those resources would need to be defined beforehand.

Generalising this is currently not possible, since there are no "wildcard" resources (as there are wildcard currency limits -we use those as well). I think that would be even harder to implement?

What I'd like is to let users define their own policies, so they could say that, for instance, different batches of jobs, even from the same user, should run together as much as possible... or the other way round... We have multiple use cases here, and want to remain as flexible as possible.

In fact, this can even be implemented with the current system, but in a very convoluted way... you can say (if you want only one job running at a node):

requirements = slot1_1_Tag =!= "MyTag" && slot1_2_Tag =!= "MyTag" && slot1_3_Tag =!= "MyTag"...

However, we have some 64-core nodes (and I think that's not so uncommon nowadays), so this is just a *really* cumbersome way of doing it, at least compared to:

requirements = !Member("MyTag",ChildTag)

That's why I think this syntax would just expose some functionality that is already there, but in an much more convenient way, and also more consistent with the new ClassAd syntax.

Best,

Joan

On 02/04/2016 07:33 PM, Greg Thain wrote:
On 02/04/2016 07:34 AM, Joan Piles wrote:

If you pair it with STARTD_JOB_ATTRS, it would allow you to make cool
things like only allowing one job of a type to run in one machine:


Note that with custom machine resources, defined with
MACHINE_RESOURCE..., HTcondor can do this in a first-class way. GPUs are
a classical example of this kind of resource, but other resources can be
allocated in the same ways.

-greg
_______________________________________________
HTCondor-devel mailing list
HTCondor-devel@xxxxxxxxxxx
https://lists.cs.wisc.edu/mailman/listinfo/htcondor-devel

--
Dr. Joan Piles
ZWE Scientific Computing
Max Planck Institute for Intelligent Systems
(p) +49 7071 601 1750

Attachment: smime.p7s
Description: S/MIME Cryptographic Signature

[← Prev in Thread] Current Thread [Next in Thread→]