incubator-s4-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Karthik Kambatla <kkamb...@cs.purdue.edu>
Subject Re: [Discuss] Runtime Isolation for s4?
Date Thu, 31 May 2012 13:46:29 GMT
Hi Nan

I agree with you that PEs should run in isolation, and have fair resource
allocation. Though they are very much inter-dependent, I believe these are
two different issues we are trying to solve.

PE Isolation:

   1. For complete isolation of each PE, we could run each PE in a
   different JVM. However, this comes at the expense of having an insane
   number of PEs running on each machine in the cluster.
   2. To avoid the large number of PEs possible by following 1, we can run
   all PEs corresponding to one application in a different JVM. If one of the
   PEs does crash the JVM, only PEs of the same application are affected.

PE Resource allocation: I would categorize your example of frequent GC
calls under this, where one PE is monopolizing the resources due to its
application logic.

   1. If adopt method 2 for isolation, it is the application logic that is
   favoring one PE over the other and I think it is alright.
   2. Dynamic load balancing (Leo's suggestion): By moving the deprived PE
   away from the monopolizing PE, we increase the chances of fair allocation.

Thoughts??

Thanks
Karthik

On Thu, May 31, 2012 at 2:42 AM, Nan N. Jiang <jiangnan@yahoo-inc.com>wrote:

> Hi Flavio and Leo,
>
> We are involving more and more aspects here in this topic.
> That will make it more confused.
> Let’s keep it sample and focus on my first question: runtime isolation.
> After that, we can talk about other aspect.
>
> For Runtime isolation,
> One application could kill another is just one example.
> Here, if we separate different in different jvms, then the unit of
> isolation is still application level.
> Actually, it’s possible that one PE could affect other PEs.
> For example:
> PE A has a lot of traffic, then PE A may trigger Java GC very frequently.
> Then other PEs in the same JVM will be affected.
> This is only one scenario that one PE will affect other PE.
> There should be more because all these PEs are sharing the same env.
> So, why won’t we think about to have a PE level isolation?
>
> About the dynamic load balancing suggested by Leo, I think it’s more about
> how to do the partition.
> So this should not satisfy my use case.
>
> Regards,
> -Nan
>
>
>
>
> On 5/30/12 5:33 AM, "Flavio Junqueira" <fpj@yahoo-inc.com> wrote:
>
> Check the excerpt from Nan's message down the thread.The main concern
> seems to be the consequences of sharing a jvm. It's unclear to me if load
> balancing alone would solve all the issues he is bringing up, although
> resource utilization seems to be an important issue. It called my attention
> the comment about an application killing another in the same jvm, which I'm
> not sure we can prevent in a practical manner. I'm particularly concerned
> about having to react aggressively and causing instabilities.
>
> Nan, Could you comment, please? if we provide dynamic load balancing as
> Leo suggests, would it satisfy your use cases?
>
> -Flavio
>
>
> On May 29, 2012, at 11:12 PM, Leo Neumeyer wrote:
>
> when starting s4 service, all applications deployed will run in the
> same JVM.
> This can be a disaster.
> Different application will fight for the JVM/hardware resource.
> Bad application even could kill other applications.
>
> Yes. Application Isolation could release the pain.
> But all applications are still fight for the resource in the same node.
> And all the PE in the same Application still fight for JVM resource.
>
>
>
>

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