Pradeep, by two level scheduling is meant that the decision on how to use a certain resource offer is delegated to the framework. It's a framework's scheduler that decides whether to make use of the incoming offer or to reject it and wait for another "more suitable" one based on the resource type it's offered and / or attributes associated with the offered slave.
On Thu, Feb 5, 2015 at 3:02 PM, Pradeep Kiruvale <[email protected]> wrote: > Hi Timothy, > > Thanks for your reply. > > Ya, even I also would like to consider all the dimensions while scheduling > an application, but here I just mentioned one dimension. > > But I did not understand what do you mean by the two level scheduler? You > mean some scheduling decisions happen > in master level and some in slave level? > > And how these decisions are made? is it uses best fit algorithm at master? > > Regards, > Pradeep > > On 5 February 2015 at 14:55, Timothy Chen <[email protected]> wrote: > >> Hi Pradeep, >> >> First of all I think the notion of optimal is not just a single dimension >> of being task duration, but also considering lots of other dimensions such >> as throughput, fairness, latency, SLA and more. >> >> Mesos is a two level scheduler, which means it's not doing all the >> scheduling at a single point (master), but instead cooperate with >> frameworks to have a good scheduling decision. >> >> So Mesos can achieve it with multiple attributes or resources as you >> mentioned with the help of frameworks. >> >> Tim >> >> On Feb 5, 2015, at 9:09 PM, Dario Rexin <[email protected]> wrote: >> >> Hi Pradeep, >> >> I am actually working on a patch for ARM support. I already have Mesos >> running on ARMv7, just need to polish it a bit and I still have 1 failing >> test. Expect news about this soon. >> >> Cheers, >> Dario >> >> On Feb 5, 2015, at 1:46 PM, Pradeep Kiruvale <[email protected]> >> wrote: >> >> Hi Dario, >> >> Thanks for the reply and clarification. >> >> How hard is to port to ARM? is there lot of architecture related code? >> Any idea? >> >> Regards, >> Pradeep >> >> On 5 February 2015 at 12:01, Dario Rexin <[email protected]> wrote: >> >>> There is currently no support for ARM cpus. GPUs and FPGAs could be >>> added to the resources in the future but are also not supported yet. >>> Scheduling tasks on machines that have a specific configuration (powerful >>> GPU or sth like that) can be done with attributes. There's however no way >>> to isolate those resources like we do with CPU and RAM. >>> >>> >>> >>> > On 05.02.2015, at 11:10, Chengwei Yang <[email protected]> >>> wrote: >>> > >>> >> On Thu, Feb 05, 2015 at 12:00:28AM +0100, Pradeep Kiruvale wrote: >>> >> Hi All, >>> >> >>> >> I am new to Mesos and I have heard and read lot about it. >>> >> >>> >> I have few doubts regarding the resource allocation by the mesos, >>> please help >>> >> me >>> >> to clarify my doubts. >>> >> >>> >> In a data center, if there are thousands of heterogeneous nodes >>> >> (x86,arm,gpu,fpgas) then is the mesos can really allocate a co-located >>> > >>> > First, does mesos can run on arm, gpu, fpga? >>> > >>> > Seconds, does your tasks run on all archs? >>> > >>> > -- >>> > Thanks, >>> > Chengwei >>> > >>> >> resources for any incoming application to finish the task faster? >>> >> >>> >> How these resource constraints are solved? what kind of a constraint >>> solver it >>> >> uses? >>> >> >>> >> Is the policy maker configurable? >>> >> >>> >> Thanks & Regards, >>> >> Pradeep >>> > >>> >> >> >> >
