[Simgrid-user] Regarding task simulation

M Vivekananda Reddy m_reddy at da-iict.org
Tue Jun 7 05:26:11 CEST 2005


 Arnaud Legrand said:
> Hi,
> Yesterday Vivekananda Reddy wrote:
>> I am using MSG. I want to model an application(like MPI) which
>> has
>> 'n' number of tasks which communicate with each other.
> Well, I've been using a lot MPI in the past and there is no notion
> of task in MPI so I hardly see what you mean... In MPI, there are
> communicating process running (most of the time) a SPMD code. So
> if  you know how  to write  your code with MPI, you should not
> have any  problem for writing it with MSG.

 By saying like MPI I was treating a task as an MPI process and "one
 task communicates with other related tasks", I meant it to be
 treated as "one MPI process sending and receiving messages from
 other MPI processes". I wanted to know whether a task in simgrid can
 perform both computation and communication.

>> The amount of work that a task has to be done can be provided in
>> Mflops i.e, the computaion size of the task. But I am not able to
>> find out how to attach to the task, the amount of communication
>> it
>> has to do with others.
> The scheduling literature is prolific about scheduling models so I
> really don't know what you're talking about here. Could you
> precise  what model you have in mind? If you are talking about a
> parallel      task model, where a task would be defined by a set
> of processors, an  amount of computations to be performed on each
> processor, and an      amount of communications to be performed
> all along the computations (e.g. data communications in a very
> simple pattern that induce synchronization), such a model can be
>    easily implemented with SURF (the simulation kernel on which
> MSG is  now based). More and more people have been asking about
> such a model  but I haven't been working on it so far because:
>    1) other coding issues in SURF/MSG/GRAS were more important
> than
> this one
>    2) a clear and formal model still had to be defined
>    3) an MSG API also had to be defined (and what people had in
> mind
> was not always clear)
>    4) I didn't have the time nor the need for it yet ;)
>> There is an attribute under the master tag in the example
>> deployment
>> file for the MSG commented as "communication size of tasks". But
>> I
>> am not sure whether it is the communication the task does with
>> other
>> or the communication cost required to move the task to a slave
>> machine form a master.
> This is the communication cost required to move the task to a
> slave
> machine from the master. In MSG when you create a task, you have
> to  precise its computation size (the amount of Mflops needed to
> process  this task when you use the MSG_task_execute function)
> and its         communication size (the amount of data that are
> attached to this      task and should therefore be communicated
> when you move this task).
>> Please clarify regarding this.
> I hope that now, everything has been clarified...


> Best regards.
>      Arnaud Legrand
> _______________________________________________
> Simgrid-user mailing list
> Simgrid-user at lists.gforge.inria.fr
> http://lists.gforge.inria.fr/cgi-bin/mailman/listinfo/simgrid-user

More information about the Simgrid-user mailing list