Re: [Gems-users] Message size and network latency

[Alberto Ros Bardisa <a.ros@xxxxxxxxxxx>]

On Wed, 2007-01-31 at 12:01 -0600, Mike Marty wrote:
> Also you might just consider changing Throttle.C to divide the size of the
> message by the width of the link.  Add this result to the fixed enqueue
> latency.  This approximation might serve your needs.
> 
> --Mike
> 

Thank you, I'll try to do this change and measure its effect.

Alberto Ros.

> > This is a limitation of the simple networking code.  The MessageSize
> > certainly affects the bandwidth of a link and therefore can cause
> > subsequent messages to incur queuing delay.  However if the size of a
> > message exceeds the size of a link, then it doesn't incur additional
> > queuing delay for _that_ message.
> >
> > It is definitely an area of GEMS that needs improvement I think.  On the
> > other hand, usually the only messages that exceed the link size are data
> > messages, and it is reasonable to assume that the critical word arrives
> > first?
> >
> > Thus the latency of a message is affected by queuing delays and the fixed
> > latencies of when a message is enqueued to the next link on the
> > interconnect.  These fixed latencies are specified in the Slicc file when
> > injecting the message into the network and the fixed latencies specified
> > for each link in the interconnect.  The latter are either specified for
> > each link using a FILE_SPECIFIED network, or the NETWORK_LINK_LATENCY and
> > ONCHIP_LINK_LATENCY parameters for the auto-generated topologies.
> >
> > You might want to e-mail Valentin Puente at the University of Cantabria.
> > At one time he integrated his more-detailed interconnect simulator into
> > Ruby.  Unfortunately for me, I cannot read Spanish code and comments, but
> > this shouldn't be a problem for you ;-)
> >
> > --Mike
> >
> >
> > > Dear list,
> > >
> > > I have noticed that changing the message size of some messages does
> > > not affect the network latency for that message.
> > >
> > > For example in the MOESI_CMP_directory coherence protocol, I have
> > > changed the message size from "MessageSizeType:ResponseLocal_Data" to
> > > "MessageSizeType:Response_Control", and I have measured the latency of
> > > the message before and after the change, finding the same result (8
> > > cycles).
> > >
> > > How does the MessageSize field affect the behaviour of the network?
> > > What other parameters affects network latency?
> > >
> > > Thanks in advance,
> > >
> > > Alberto Ros.
> > >
> > >
> > > _______________________________________________
> > > Gems-users mailing list
> > > Gems-users@xxxxxxxxxxx
> > > https://lists.cs.wisc.edu/mailman/listinfo/gems-users
> > > Use Google to search the GEMS Users mailing list by adding "site:https://lists.cs.wisc.edu/archive/gems-users/"; to your search.
> > >
> > _______________________________________________
> > Gems-users mailing list
> > Gems-users@xxxxxxxxxxx
> > https://lists.cs.wisc.edu/mailman/listinfo/gems-users
> > Use Google to search the GEMS Users mailing list by adding "site:https://lists.cs.wisc.edu/archive/gems-users/"; to your search.
> >
> _______________________________________________
> Gems-users mailing list
> Gems-users@xxxxxxxxxxx
> https://lists.cs.wisc.edu/mailman/listinfo/gems-users
> Use Google to search the GEMS Users mailing list by adding "site:https://lists.cs.wisc.edu/archive/gems-users/"; to your search.