[openstack-dev] [nova][scheduler] Proposal: FairShareScheduler.

Sylvain Bauza sbauza at redhat.com
Thu Jul 10 14:01:04 UTC 2014


Hi Lisa,

Le 10/07/2014 15:40, Lisa a écrit :
> Hi Joe,
>
> we know the Amazon's hot market is based on a economic model. Suppose to assign to each group a virtual money amount for renting hot instances.
> So, the money may reflect the "share" concept in our model. This means that a big money amount may rent potentially more resources (actually it depends by the current resource price, the offer, etc).
> As in the real World, this economic model emphasizes the difference between rich and poor. So a rich group may rent all resources available for long time.
> For sure this approach maximizes the resource utilization but it is not so fair. It should guarantees that the usage of the resources is fairly distributed among users and teams according to their mean number of VMs they have got by considering the portion of the resources allocated to them (i.e. share) and the evaluation of the effective resource usage consumed in the recent past. 
> Probably this issue may be solved by defining fair lease contracts between the user and the resource provider. What  do you think?
> To make it more clear, for simplicity let's say that we have only two teams, "A" and "B" who are undertaking activities of Type3. The site administrator assigns 1000$ to the group "A" and just 100$ to "B". Suppose "A" and "B" need both more resources at the same time. So, for sure "B" will be able to rent something only when "A" releases its resources or when "A" becomes poor. Instead, with a fair-share algorithm, "B" would be able to rent some resources because the purchasing power of "A" is adjusted by the recent trading.
> At the moment I'm still not really sure that this approach can cover this use case.

The idea behind what you mention is that there are possibly different
policies for overcommitting the resources, but the contract remains the
same : by accepting a Type 3 lease, users from either group A or group B
accepts that an external event can terminate its lease.

>From what is coming the event, and how the event is triggered is a
question of policy to me. Ideally, it should be operator-driven thanks
to a choice in the policies he wants to play with.

In any way, that doesn't change how the workflow is made, nor the
contract (ie. the API, as Joe said).

-Sylvain


> thanks a lot.
> Cheers,
> Lisa
>
>
>
>
> On Tue, Jul 8, 2014 at 4:18 AM, Lisa <lisa.zangrando at pd.infn.it <http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev>> wrote:
>
> >/  Hi Sylvain,
> />/
> />/
> />/ On 08/07/2014 09:29, Sylvain Bauza wrote:
> />/
> />/ Le 08/07/2014 00:35, Joe Gordon a écrit :
> />/
> />/
> />/ On Jul 7, 2014 9:50 AM, "Lisa" <lisa.zangrando at pd.infn.it <http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev>> wrote:
> />/ >
> />/ > Hi all,
> />/ >
> />/ > during the last IRC meeting, for better understanding our proposal (i.e
> />/ the FairShareScheduler), you suggested us to provide (for the tomorrow
> />/ meeting) a document which fully describes our use cases. Such document is
> />/ attached to this e-mail.
> />/ > Any comment and feedback is welcome.
> />/
> />/ The attached document was very helpful, than you.
> />/
> />/ It sounds like Amazon's concept of spot instances ( as a user facing
> />/ abstraction) would solve your use case in its entirety. I see spot
> />/ instances as the general solution to the question of how to keep a cloud at
> />/ full utilization. If so then perhaps we can refocus this discussion on the
> />/ best way for Openstack to support Amazon style spot instances.
> />/
> />/
> />/
> />/
> />/ Can't agree more. Thanks Lisa for your use-cases, really helpful for
> />/ understand your concerns which are really HPC-based. If we want to
> />/ translate what you call Type 3 in a non-HPC world where users could compete
> />/ for a resource, spot instances model is coming to me as a clear model.
> />/
> />/
> />/ our model is similar to the Amazon's spot instances model because both try
> />/ to maximize the resource utilization. The main difference is the mechanism
> />/ used for assigning resources to the users (the user's offer in terms of
> />/ money vs the user's share). They differ even on how they release the
> />/ allocated resources. In our model, the user, whenever requires the creation
> />/ of a Type 3 VM, she has to select one of the possible types of "life time"
> />/ (short = 4 hours, medium = 24 hours, long = 48 hours). When the time
> />/ expires, the VM is automatically released (if not explicitly released by
> />/ the user).
> />/ Instead, in Amazon, the spot instance is released whenever the spot price
> />/ rises.
> />/
> />/
> /I think you can adapt your model your use case to the spot instance model
> by allocating different groups 'money' instead of a pre-defined share. If
> one user tries to use more then there share they will run out of 'money.'
>  Would that fully align the two models?
>
> Also why pre-define the different life times for type 3 instances?
>
>
>  >/
> />/
> />/
> />/ I can see that you mention Blazar in your paper, and I appreciate this.
> />/ Climate (because that's the former and better known name) has been kick-off
> />/ because of such a rationale that you mention : we need to define a contract
> />/ (call it SLA if you wish) in between the user and the platform.
> />/ And you probably missed it, because I was probably unclear when we
> />/ discussed, but the final goal for Climate is *not* to have a start_date and
> />/ an end_date, but just *provide a contract in between the user and the
> />/ platform* (see
> />/ https://wiki.openstack.org/wiki/Blazar#Lease_types_.28concepts.29 )
> />/
> />/ Defining spot instances in OpenStack is a running question, each time
> />/ discussed when we presented Climate (now Blazar) at the Summits : what is
> />/ Climate? Is it something planning to provide spot instances ? Can Climate
> />/ provide spot instances ?
> />/
> />/ I'm not saying that Climate (now Blazar) would be the only project
> />/ involved for managing spot instances. By looking at a draft a couple of
> />/ months before, I thought that this scenario would possibly involve Climate
> />/ for best-effort leases (see again the Lease concepts in the wiki above),
> />/ but also the Nova scheduler (for accounting the lease requests) and
> />/ probably Ceilometer (for the auditing and metering side).
> />/
> />/ Blazar is now in a turn where we're missing contributors because we are a
> />/ Stackforge project, so we work with a minimal bandwidth and we don't have
> />/ time for implementing best-effort leases but maybe that's something we
> />/ could discuss. If you're willing to contribute to an Openstack-style
> />/ project, I'm personnally thinking Blazar is a good one because of its
> />/ little complexity as of now.
> />/
> />/
> />/
> />/ Just few questions. We read your use cases and it seems you had some
> />/ issues with the quota handling. How did you solved it?
> />/ About the Blazar's architecture (
> />/ https://wiki.openstack.org/w/images/c/cb/Climate_architecture.png): the
> />/ resource plug-in interacts even with the nova-scheduler?
> />/ Such scheduler has been (or will be) extended for supporting the Blazar's
> />/ requests?
> />/ Which relationship there is between nova-scheduler and Gantt?
> />/
> />/ It would be nice to discuss with you in details.
> />/ Thanks a lot for your feedback.
> />/ Cheers,
> />/ Lisa
> />/
> />/
> />
>
>
> _______________________________________________
> OpenStack-dev mailing list
> OpenStack-dev at lists.openstack.org
> http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.openstack.org/pipermail/openstack-dev/attachments/20140710/a62c8214/attachment.html>


More information about the OpenStack-dev mailing list