mesos-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Anshuman Goswami <goswami.anshu...@gmail.com>
Subject Re: GPGPU SLA support
Date Mon, 11 Jan 2016 19:58:32 GMT
Thanks for the feedback.

I wanted to know if the mechanism would at all be acceptable to the
community. It involves adapting GPU kernels to incorporate timeslicing
functionality. Also, there is a limitation if GPU kernels without the
source or ptx are being used (e.g., close-sourced libraries like cuDNN,
cuFFT etc.).

On Mon, Jan 11, 2016 at 2:48 PM, Vaibhav Khanduja <vaibhavkhanduja@gmail.com
> wrote:

> What is your exact proposal?
>
> Having to use GPU as a sharable resource is definately beneficial. It would
> really makes sense, for any framework to control and schedule jobs, asking
> for GPU resources similar to sharable resources.
>
> Do you have a proposal of changes required (if) in Mesos?
>
> On Mon, Jan 11, 2016 at 11:34 AM, Anshuman Goswami <
> goswami.anshuman@gmail.com> wrote:
>
> > Any thoughts?
> >
> > On Mon, Jan 4, 2016 at 2:56 PM, Anshuman Goswami <
> > goswami.anshuman@gmail.com
> > > wrote:
> >
> > > I have posted this on the user list. Reposting here for continuity of
> > this
> > > thread.
> > >
> > > I have been working on GPU QoS by modifying the GPU kernels. Something
> > > along the lines of this
> > > <http://drona.csa.iisc.ernet.in/~mjt/papers/asplos2013.pdf>. The basic
> > > mechanism is to time slice a grid by running a subset of thread blocks
> in
> > > one invocation. The main challenge is in controlling the subset. The
> > > modification cost is paid only once during the first invocation of the
> > > kernel.
> > >
> > > Would this be usable?
> > >
> > > On Mon, Mar 23, 2015 at 12:37 PM, Niklas Nielsen <niklas@mesosphere.io
> >
> > > wrote:
> > >
> > >> Should be available now through http://mesos.apache.org/community/
> > >>
> > >> Niklas
> > >>
> > >> On 23 March 2015 at 12:24, Anshuman Goswami <
> goswami.anshuman@gmail.com
> > >
> > >> wrote:
> > >>
> > >> > Thanks!
> > >> >
> > >> > On Mon, Mar 23, 2015 at 3:18 PM, Niklas Nielsen <
> niklas@mesosphere.io
> > >
> > >> > wrote:
> > >> >
> > >> > > By sending an email with the topic 'Subscribe' to
> > >> > > modules-subscribe@mesos.apache.org
> > >> > > I apologize for having this on the website; will get it up now.
> > >> > >
> > >> > > Niklas
> > >> > >
> > >> > > On 20 March 2015 at 19:49, Anshuman Goswami <
> > >> goswami.anshuman@gmail.com>
> > >> > > wrote:
> > >> > >
> > >> > > > How do I subscribe to the 'modules' mailing list?
> > >> > > >
> > >> > > > On Fri, Mar 20, 2015 at 10:16 PM, Anshuman Goswami <
> > >> > > > goswami.anshuman@gmail.com> wrote:
> > >> > > >
> > >> > > > > Thanks Niklas for the pointer...will get back as I
dig in
> > more....
> > >> > > > >
> > >> > > > > On Fri, Mar 20, 2015 at 6:20 PM, Niklas Nielsen <
> > >> > niklas@mesosphere.io>
> > >> > > > > wrote:
> > >> > > > >
> > >> > > > >> Sounds like you are on the right track; let us
know if you
> have
> > >> any
> > >> > > > >> problems developing/running modules.
> > >> > > > >> We created a modules mailing list too where we
can go into
> more
> > >> > > details
> > >> > > > -
> > >> > > > >> modules@mesos.apache.org
> > >> > > > >>
> > >> > > > >> Cheers,
> > >> > > > >> Niklas
> > >> > > > >>
> > >> > > > >> On 20 March 2015 at 12:47, Anshuman Goswami <
> > >> > > goswami.anshuman@gmail.com
> > >> > > > >
> > >> > > > >> wrote:
> > >> > > > >>
> > >> > > > >> > Hi mesos developers,
> > >> > > > >> >
> > >> > > > >> > I am just getting started with understanding
the
> internals. I
> > >> am
> > >> > > > >> working on
> > >> > > > >> > a project to add some SLA support to GPGPU
resources and
> want
> > >> to
> > >> > > have
> > >> > > > >> that
> > >> > > > >> > work under mesos.
> > >> > > > >> >
> > >> > > > >> > After reading this Q&A (
> > >> > > > >> >
> > >> > > > >> >
> > >> > > > >>
> > >> > > >
> > >> > >
> > >> >
> > >>
> >
> http://stackoverflow.com/questions/27872558/does-apache-mesos-recognize-gpu-cores
> > >> > > > >> > ),
> > >> > > > >> > I started looking into the pluggable isolator
modules. It
> > seems
> > >> > like
> > >> > > > the
> > >> > > > >> > right approach until the feature works, is
tested and
> becomes
> > >> > > stable.
> > >> > > > >> >
> > >> > > > >> > Any feedback on whether this makes sense?
or something like
> > >> this
> > >> > > > already
> > >> > > > >> > exists?
> > >> > > > >> >
> > >> > > > >> > Thanks,
> > >> > > > >> > Anshuman
> > >> > > > >> >
> > >> > > > >>
> > >> > > > >
> > >> > > > >
> > >> > > >
> > >> > >
> > >> >
> > >>
> > >
> > >
> >
>

Mime
  • Unnamed multipart/alternative (inline, None, 0 bytes)
View raw message