Re: MPI/RT for Myrinet

On 7/29/05, Dan Strahs <[email protected]> wrote:
>=20
> It's my impression from reading that you will need a real-time kernel
> (that is, if you're using Linux/UNIX/Windows/Mac/etc). Without that basic=
,
> there is no support for realtime in GM - you must have the kernel
> interrupts to pass messages in realtime.

I'm using RTLinux to get time guarantees needed by this MPI
application. But MPICH-GM itself doesn't support deterministic
communication and is likely to blow the hard real-time guarantees
provided by RTLinux. So I was wondering if I could change this
behavior to provide time guarantees?

Also, what role does the interrupt handler(in GM linux driver code)
have to play when MPI_Recv is invoked with polling mode to receive a
msg? Does the channel interface viz. ch_gm support blocking
send/receive operations?

>=20
> There is support for these kernels - you'll have to investigate your
> options.
>=20
> Dan Strahs
>=20
> On Thu, 28 Jul 2005, Patrick Geoffray wrote:
>=20
> > Ashish Bijlani wrote:
> > > On 7/28/05, Patrick Geoffray <[email protected]> wrote:
> > >
> > >>Ashish Bijlani wrote:
> > >>
> > >>>Hi People,
> > >>>
> > >>>I've a setup of 8 nodes, each having M3F-PCIXD-2 interface. I want t=
o
> > >>>pass messages in Real-Time using MPI. Does this Myrinet interface
> > >>>support Real-Time MPI(MPI/RT)?
> > >>
> > >>No. MPI/RT never really took off. I don't know anybody using it.
> > >>
> > >
> > >
> > > Can I use MPICH-GM in Real-Time to pass messages to other nodes? If
> > > not, then what are the bottlenecks involved?
> >
> > You can use MPICH-GM to pass messages, but you won't have hard real tim=
e
> > guarantees. First, you need a real time OS, and then you need to have
> > policies and admission control and scheduling in the NIC or in the lib
> > to guarantee the RT message deliveries. A lot of work.
> >
> > >>>If not, then how can I pass messages in
> > >>>Real-Time across nodes?
> >
> > If your RT deadlines fits in the current best effort performance, you
> > have nothing to do, but you have no guarantees either.
> >
> > Patrick
> >
>=20
>