That's correct (for now). We're looking into features that would support preemption of running tasks, but currently a user/admin would have to manually kill long-running tasks to scale down an over-provisioned framework. Marathon also has a nice API (web or REST) for scaling down the number of instances of a long-running service.
On Mon, Aug 18, 2014 at 1:43 AM, Jun Feng Liu <[email protected]> wrote: > Thanks, Adam.. Sounds like it is going to be pretty effective when all the > framework running a short tasks, then mesos can balance the resource > allocation based on the DRF among the framework quickly. If one framework > is happening to run some long tasks and take too many resources, mesos have > to wait until some resource being free up to assign to other framework. Is > it correct? > > Best Regards > > > *Jun Feng Liu* > IBM China Systems & Technology Laboratory in Beijing > > ------------------------------ > [image: 2D barcode - encoded with contact information] > *Phone: *86-10-82452683 > * E-mail:* *[email protected]* <[email protected]> > [image: IBM] > > BLD 28,ZGC Software Park > No.8 Rd.Dong Bei Wang West, Dist.Haidian Beijing 100193 > China > > > > > > *Adam Bordelon <[email protected] <[email protected]>>* > > 2014/08/18 16:26 > Please respond to > [email protected] > > To > "[email protected]" <[email protected]>, > cc > Jay Buffington <[email protected]> > Subject > Re: mesos scheduling > > > > > Mesos uses a fair-sharing algorithm[1] to ensure that each framework > registered with Mesos is ensured its fair share of resources. If you want > more control over the groupings and weights of different frameworks, check > out the roles and weights parameters: mesos-master --roles="services,batch" > and --weights="services=2,batch=1" as described at > *http://mesosphere.io/docs/mesos/deep-dive/mesos-master/* > <http://mesosphere.io/docs/mesos/deep-dive/mesos-master/> > > Mesos uses these algorithms and parameters to decide which framework gets > the next offer, so it won't affect already running tasks if one framework > is already hogging the cluster when you start a new framework. But if you > start killing tasks from the over-provisioned framework, those resources > will be offered to the new framework(s) until it reaches its fair share. > > [1] *http://static.usenix.org/event/nsdi11/tech/full_papers/Ghodsi.pdf* > <http://static.usenix.org/event/nsdi11/tech/full_papers/Ghodsi.pdf> > > > On Sun, Aug 17, 2014 at 7:06 PM, Jun Feng Liu <*[email protected]* > <[email protected]>> wrote: > Thanks Jay.. Dose it mean if one of scheduler/frame need a lot resource > and keep ask for more resources from mesos, then it will cause other > framework/scheduler hard to get resources? Any way I can configure the > mesos to setup a resource consuming boundary for each framework? > Best Regards > > > * Jun Feng Liu* > IBM China Systems & Technology Laboratory in Beijing > > ------------------------------ > *Phone: *86-10-82452683 > * E-mail:* *[email protected]* <[email protected]> > [image: IBM] > > BLD 28,ZGC Software Park > No.8 Rd.Dong Bei Wang West, Dist.Haidian Beijing 100193 > China > > > > > *Jay Buffington <**[email protected]* <[email protected]>*>* > Sent by: *[email protected]* <[email protected]> > > 2014/08/18 02:44 > > > Please respond to > *[email protected]* <[email protected]> > > To > *[email protected]* <[email protected]>, > cc > Subject > Re: mesos scheduling > > > > > > > On Sun, Aug 17, 2014 at 6:13 AM, Jun Feng Liu <*[email protected]* > <[email protected]>> wrote: > I am trying to better understand how mesos allocator works. In the offer > resource model, will mesos send the same offer to multiple framework? Or it > just send all resource to one framework then wait the response from the the > framework then try the next one? > > Mesos sends an offer to one scheduler (a scheduler is part of a framework) > at a time. That scheduler will have the offer until it uses it, gives it > back or mesos rescinds it. > > This strategy was referred to as "pessimistic" by Google's Omega paper [1] > and has drawbacks. In order to address these points a new type of offer, > an Optimistic Offer, is being considered. See > *https://issues.apache.org/jira/browse/MESOS-1607* > <https://issues.apache.org/jira/browse/MESOS-1607> > > Jay > > [1] > *http://eurosys2013.tudos.org/wp-content/uploads/2013/paper/Schwarzkopf.pdf* > <http://eurosys2013.tudos.org/wp-content/uploads/2013/paper/Schwarzkopf.pdf> > >
