On Thu, Sep 27, 2012 at 1:06 PM, Martin Packman <span dir="ltr"><<a href="mailto:martin.packman@canonical.com" target="_blank">martin.packman@canonical.com</a>></span> wrote:<br><div class="gmail_quote"><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">
What do juju users want from a cpu constraint on instances?<br>
<br>
I'm adding support in MaaS, and have some choices on what the number<br>
means exactly.<br>
<br>
<<a href="http://docs.openstack.org/api/openstack-compute/2/content/Get_Flavor_Details-d1e4317.html" target="_blank">http://docs.openstack.org/api/openstack-compute/2/content/Get_Flavor_Details-d1e4317.html</a>><br>
<br>
For OpenStack, 'vcpus' is used, which (though not clearly documented?)<br>
is the number of cores the instance has.<br>
<br>
<<a href="http://aws.amazon.com/ec2/instance-types/" target="_blank">http://aws.amazon.com/ec2/instance-types/</a>><br>
<br>
In EC2 abstract 'compute units' are used, which they correspond<br>
approximately to 1.0-1.2 GHz processors from 2007. However, the<br>
documentation pushes you towards benchmarking on different instance<br>
types over than treating these numbers too seriously.<br>
<br>
So, either presenting the number of cpus, or the total GHz, seem like<br>
reasonable options. Which would you find more useful in general? Using<br>
tags in MaaS can constrain on more precise details regardless.<br>
<span class="HOEnZb"><font color="#888888"><br></font></span></blockquote><div><br></div><div>Its an interesting question, previously when constraints were being formulated, one idea was that we'd have a simple benchmark program that could give some approximation of a notional linear scale roughly corresponding to an ec2 unit given a cpu intensive workload, except via the benchmark in a cross provider fashion. This a fairly naive approach that occludes alot of the variations across architectures and versions there of and workloads as well, but is nice for its simplicity. Ie. python -m test.pystone with a different normalization. For now (ie delivering generic constraints for 12.10) i would suggest pursuing the even more naive approach of openstack and just use cpu core count. additional facilities around gpu compute facilities, processor family, etc would be nice to expose as a base set of automatic maas tags.</div>
<div><br></div><div>cheers,</div><div><br></div><div>Kapil</div><div><br></div></div>