Open MPI logo

Open MPI Development Mailing List Archives

  |   Home   |   Support   |   FAQ   |   all Development mailing list

Subject: Re: [OMPI devel] iWARP development
From: Ralph Castain (rhc_at_[hidden])
Date: 2013-12-12 17:44:18


We already have conditional compilation there at configure time (see
ompi/mca/btl/sctp/configure.m4) so it won't built unless requested.
Likewise, you can use the MCA params to select to use it or not (-mca btl
sctp,sm,self).

I'm not aware of any sctp-specific tests. However, there are lots of simple
MPI tests you could run that would test it as a BTL - anything that would
exchange data across nodes, for example.

Nobody has spoken up to "own" the sctp btl at this time, so I think you'd
have a pretty free hand with it (subject to our usual critique to ensure
consistency across the BTLs).

On Thu, Dec 12, 2013 at 4:31 PM, Prindeville, Philip <
philip.prindeville_at_[hidden]> wrote:

> I think I understand.
>
>
>
> I’ll pull a copy of trunk and dig around in there.
>
>
>
> Is there a reason that the code can’t be gated by conditional compilation
> flags or detect the environment at run-time?
>
>
>
> Is there anything in the way of a set of verification tests? And what’s
> the provenance of the SCTP code that’s in trunk?
>
>
>
> Thanks,
>
>
>
> -Philip
>
>
>
>
>
> *From:* devel [mailto:devel-bounces_at_[hidden]] *On Behalf Of *Ralph
> Castain
> *Sent:* Thursday, December 12, 2013 8:41 AM
> *To:* Open MPI Developers
>
> *Subject:* Re: [OMPI devel] iWARP development
>
>
>
> Be aware that we removed SCTP from the 1.7 release series because of its
> unknown state of repair - I don't believe anyone has tested it in quite
> some time, and nobody has been maintaining it so far as we know. Not saying
> it won't work - just saying that I don't think we know :-/
>
>
>
> On Wed, Dec 11, 2013 at 6:07 PM, Jeff Squyres (jsquyres) <
> jsquyres_at_[hidden]> wrote:
>
> On Dec 11, 2013, at 5:33 PM, "Prindeville, Philip" <
> philip.prindeville_at_[hidden]> wrote:
>
> > I was wondering what the current state of iWARP development is.
>
> Open MPI's iWARP support is part of the "openib" BTL (so named because
> OpenFabrics used to be known as OpenIB, before iWARP joined, and we never
> changed the name of our plugin after OFA became OFA).
>
>
> > There are some features we’re interested in, and from what I can tell
> the iWARP RFCs/Internet Drafts haven’t caught up to related developments.
>
> As George mentioned, we deleted the SCTP plugin from the 1.7 release
> branch -- but that's a standalone SCTP plugin, which is not what I think
> you're asking about it.
>
>
> > Part of our interest is in using SCTP as the LLP for iWARP, and updating
> that adaptation to the latest SCTP work.
>
> Gotcha.
>
>
> > For instance:
> >
> > RFC 6458 – SCTP authentication
> > RFC 6458 – SCTP out-of-order delivery
> > RFC 6458 – SCTP association end-point address changes
> > RFC 6458 – SCTP Receive Information
> > RFC 6458 – SCTP partially reliable delivery
> > RFC 6458 – SCTP key management
> > RFC 5061 – SCTP Dynamic address reconfiguration (useful for hot NIC
> swaps in a high availability environment)
> >
> > We’d also like to see load-sharing in multipath environments, and
> sender-side traffic shaping support.
>
> Sounds nifty.
>
>
> > From what I can tell, the iWARP SCTP work that’s been done predates
> RFC-6458, and hence I’m assuming it’s aligned to RFC-5043.
>
> Sure...?
>
>
> > Other questions I have:
> >
> > Has this code been tested extensively on non-x86 platforms? What about
> IA64, PPC64, ARM7, or MIPS 7K?
>
> Doubtful. The openib BTL is known to work with IB and iWARP on a variety
> of x86 platforms. I have no idea, and would kinda doubt, if it has been
> tested on any of the other platforms you've specified.
>
>
> > Is anyone doing any code to port SolarFlare OpenOnload stack to support
> Open MPI’s iWARP?
>
> Nope. SF has told me/others that they're happy just doing their onload
> TCP through OMPI -- they don't see the need to do their own OO plugin (but
> don't take my word for it; I certainly cannot speak for them -- feel free
> to ask them yourself).
>
>
> > And a minor note… Just looking at the 1.7.3 tarball and grepping for
> SCTP in it, I find only a few matches, such as this:
> >
> > evutil_getaddrinfo_infer_protocols(struct evutil_addrinfo *hints)
> > {
> > …
> > #ifdef IPPROTO_SCTP
> > else if (hints->ai_protocol ==
> IPPROTO_SCTP)
> > hints->ai_socktype =
> SOCK_STREAM;
> > #endif
> > }
> > }
> >
> > And this has me puzzled: SCTP is predominately a SOCK_SEQPACKET, isn’t
> it? (The whole point of using it and not TCP as a transport is it preserves
> record boundaries, supports out-of-order delivery and message interleaving,
> etc.)
> >
> > At least, that’s how it registers itself as a protocol in Linux 3.12 in
> net/sctp/protocol.c …
> >
> > Apologies if there’s a better mailing list for iWARP specific questions.
> I’m looking at a lot of this stuff for the first time and having to ramp up
> quickly.
> >
> > Thanks,
> >
> > -Philip
> >
> >
>
> > _______________________________________________
> > devel mailing list
> > devel_at_[hidden]
> > http://www.open-mpi.org/mailman/listinfo.cgi/devel
>
> --
> Jeff Squyres
> jsquyres_at_[hidden]
> For corporate legal information go to:
> http://www.cisco.com/web/about/doing_business/legal/cri/
>
>
> _______________________________________________
> devel mailing list
> devel_at_[hidden]
> http://www.open-mpi.org/mailman/listinfo.cgi/devel
>
>
>
> _______________________________________________
> devel mailing list
> devel_at_[hidden]
> http://www.open-mpi.org/mailman/listinfo.cgi/devel
>