airavata-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Shameera Rathnayaka <shameerai...@gmail.com>
Subject Re: Work Stealing is not a good solution for Airavata.
Date Fri, 07 Oct 2016 16:50:27 GMT
On Fri, Oct 7, 2016 at 12:18 PM Amila Jayasekara <thejaka.amila@gmail.com>
wrote:

> You misunderstood.
>
> Let me re-phrase what mentioned earlier: (You said) "No, it is still
> valid, thread goes to thread pool doesn't say *worker* is complete that
> request, *it* is waiting until actual hpc job runs on target computer
> resoruces".
> Try to read above statement and explain to me what you refer as "worker"
> and "it" (in bold) ?
>

Worker mean, who consume from the queue (to be specific GFac). "It" means
Worker.

>
> Here is how I understood above statement. You submit a job in a thread and
> then you put that thread to "wait" state until job finishes. Is that
> correct ?
>
No, this is bad we are not putting thread to wait. we release it untill the
submitted job finish, next time GFac start to process the same request,  it
can be same thread or different thread from thread pool.


> Also, others let me know whether I read the statement incorrectly.
>
> -AJ
>
> On Fri, Oct 7, 2016 at 11:46 AM, Shameera Rathnayaka <
> shameerainfo@gmail.com> wrote:
>
> glad you got that, also don't you see that when the message comes(very
> first right arrow) and when gfac send the ack for that message(very last
> left arrow)?  That is the lifetime of a one worker queue message.
>
> On Fri, Oct 7, 2016 at 9:57 AM Amila Jayasekara <thejaka.amila@gmail.com>
> wrote:
>
> As per this diagram, it seems the thread that submits the job is not the
> same as the thread that handles output.
> At least that is what I understand.
>
> -AJ
>
> On Thu, Oct 6, 2016 at 4:26 PM, Shameera Rathnayaka <
> shameerainfo@gmail.com> wrote:
>
> Previous attachment doesn't work.
>
> On Thu, Oct 6, 2016 at 4:24 PM, Shameera Rathnayaka <
> shameerainfo@gmail.com> wrote:
>
> [image: Work Queue Message Life time.png]Hi Amila,
>
> Please find work queue message execution sequence diagram below. Hope this
> will help to understand how it works in Airavata.
>
>
>
> On Thu, Oct 6, 2016 at 4:05 PM Suresh Marru <smarru@apache.org> wrote:
>
> Just a quick top post. This is informative discussion, please continue :)
>
> I agree on that Airavata does not do Work Stealing but it implements "Work
> Queues”. Conceptually they are similar to the OS Kernel level work queens,
> but more in a distributed context -
> https://www.kernel.org/doc/Documentation/workqueue.txt
>
> Suresh
>
>
> On Oct 6, 2016, at 3:52 PM, Amila Jayasekara <thejaka.amila@gmail.com>
> wrote:
>
> On Thu, Oct 6, 2016 at 3:17 PM, Shameera Rathnayaka <
> shameerainfo@gmail.com> wrote:
>
>
>
> On Thu, Oct 6, 2016 at 2:50 PM Amila Jayasekara <thejaka.amila@gmail.com>
> wrote:
>
> On Thu, Oct 6, 2016 at 11:07 AM, Shameera Rathnayaka <
> shameerainfo@gmail.com> wrote:
>
> Hi Amila,
>
> -- Please explain how you used "work stealing" in distributed system. That
> would be interesting.
>
>
> Airavata depends on work stealing + amqp for followings,
> Fault Tolerance - This is one of major distributed system problem which
> critical in Airavata, What ever the reason experiment request processing
> shouldn't get any effect  from internal node failure. Even with the node
> failures, Airavata should be capable enough to continue experiment request
> processing or hold it until at least one node appear and then continue. How
> this is handled in Ariavata is, worker only ack for messages only after it
> completely processed it. If the node goes down without sendings  ack for
> the messages it was processing,then rabbitmq put all these un-ack messages
> back to the queue and available to consume again.
>
> Resource Utilization- Another important goal of distributed system to
> effectively use available resources in the system, namely the memory and
> processors of components.  In Airavata this will decide the throughput and
> response time of experiments. Currently, at a given time workers only get
> messages up to a preconfigured limit (the limit is prefetch count) But most
> of these jobs are async jobs. That means after worker gets fixed amount of
> jobs, it won't get any other jobs even worker capable or handling more
> jobs, waste of worker resources.
>
>
>
> You still did not answer my question. I want to know how you used "work
> stealing" in your implementation. In other words how distributed work
> stealing works in your implementation. The details  you gave above is
> unrelated and does not answer my question.
>
>
> I think I have explained, how we use work stealing (work queues). If you
> are finding a more analog solution to parallel computing work strealing
> then that is hard to explain.
>
>
> No, you have not. :-).
> Work stealing != work queues. In a distributed setting I would image
> following kind of a work stealing implementation; Every worker
> (orchestrator) maintains a request queue locally and it serve requests
> coming to the local queue. Whenever one worker runs out of more requests to
> serve it will query other distributed workers local queues to see whether
> there are requests that it can serve. If there are it can steal requests
> from other workers local queues and process. However, this model of
> computation is in efficient to do in a distributed environment. I guess
> that is the same reason we dont find much distributed work stealing
> implementations.
>
> Anyhow lets stop the discussion about work stealing now. :-)
>
>
>
>
>
>
>
>
> -- I dont see AMQP in the architecture diagram you attached above and I
> dont understand why Airavata has to depend on it. One way to figure this
> out is think about the architecture without AMQP and figure out what
> actually should happend and look for a way to do that using AMQP.
>
>
> Worker Queue is AMQP queue.
>
>
> Does the worker queue needs to be an AMQP queue ? Sorry, I dont know much
> about AMQP but it sounds like limitations you are explaining are because of
> AMQP.
>
>
> It is not, but good to use well-defined protocol instead of custom one.
> Almost all messaging systems have implemented AMQP protocol.
>
>
> Can we figure out whether others have also encountered the same/similar
> problem and how they tackled those with AMQP ? Cos the design we have is
> pretty straightforward and I believe there are systems analogous to our
> design that uses AMQP.
>
>
>
>
>
>
>
> -- Does this mean that you have a waiting thread or process within
> Airavata after submitting the job (for each work) ?
>
>
> No, once the job is submitted to the remote resource, thread goes back to
> the thread pool.
>
>
> Then, your previous explanation, (i.e., "The time needs for a worker to
> finish the work is depend on the application run time (applications runs on
> HPC machine). Theoretically, this can be from few sec to days or even
> more."), invalidates. Correct ?
>
>
> No, it is still valid, thread goes to thread pool doesn't say worker is
> complete that request, it is waiting until actual hpc job runs on target
> computer resoruces. After this hpc jobs completed then outptu data staging
> happens. After output stage to storage then it ack to the work queue
> message.
>
>
> This is confusing to me.
> Does this mean once you return thread to thread pool, it is not reusable
> for another request ? Also, how do you wait on a thread after returning it
> to the thread pool ?
> Also, why do you have to wait for HPC job to complete ? I was under the
> impression the communication is asynchronous. i.e. after job completes you
> get an email confirmation and then you start output data staging in a
> separate thread.
>
> We should probably meet and verbally discuss this.
>
> -AJ
>
>
>
> Thanks,
> Shameera.
>
>
>
>
>
> Thanks,
> Shameera.
>
>
>
> It takes more time for me to digest following right now. I will try to
> give more feedback when I properly understand them.
>
> Thanks
> -Amila
>
> That means, If a worker can read N(=prefetch count) number of messages
> from a queue without sending acknowledgment then that is the limit one
> worker can handle for given time. But most of this long-running jobs are
> asynchronous. Worker resources are free to handle more works than N. Hence
> Airavata underutilized worker resources. In the case of small jobs (small
> in runtime), this won't be a big problem.
>
> Apache Zookeeper[2] provide a way to manage distributed system components
> and most of  the latest distributed systems have been used Zookeeper to
> address all the common distributed system problems like HA, FT, Leader
> Election etc ... But in Airavata is trying to replace Rabbitmq with
> Zookeeper to achieve the same outcomes, I haven't seen any framework have
> done it. How Airavata tries to do is using Work Stealing Queues. Anyway,
> Airavata hasn't move zookeeper out of its architecture yet as it uses
> zookeeper to handle cancel requests.
>
>
>
> --
Shameera Rathnayaka

Mime
View raw message