The OpenGroup InterConnect Software Consortium has started
a working group to propose extensions to the current Unix socket API
and is seeking help from the open source community.
These extensions (ideas floating around are, but not limited to:
asynchronous paradigm to enable zero copy, I/O completion ports,
advance buffer registration to allow pinning and I/O preparation, ...)
would enable new networking technologies (Infiniband, RDMA, ...) and
would allow Unix to regain the lead over Winsock in terms of functionality
and performance.
The OpenGroup is trying to figure out how to include the open source
community in these activities. It is actively seeking the people who are
the Linux, FreeBSD, ..., thought leaders and implementors for networking
and sockets, and bring them into the working group.
Interested people should contact the working group co-chairs for details
about how to join the effort:
David Edmondson ([email protected]) or
Satya Sharma ([email protected])
The URL of the OpenGroup Interconnect Software Consortium is:
http://www.opengroup.org/icsc/
Thanks for volunteering.
+---------------------------------------------------------+
| Louis LABORDE e-mail: [email protected] |
| HP Cupertino SISL phone: (408) 447-3649 |
+---------------------------------------------------------+
> Interested people should contact the working group co-chairs for details
> about how to join the effort:
> David Edmondson ([email protected]) or
> Satya Sharma ([email protected])
>
> The URL of the OpenGroup Interconnect Software Consortium is:
> http://www.opengroup.org/icsc/
Yea, yea. The little problem here is that TOG is not getting it.
When I see a website that says "Employees of member companies should
send mail to the Consortium Administrator requesting access to the
Member area.", I sure as hell is not getting a hardon about working
on "overtaking Winsock" (and why exactly would we want to copy
broken attempts to embrace and extend? Perhaps you should talk to
a certain <i>monkey lover company</i> about it).
I would not mind to participate in something more open.
If anyone has a suggestion of a mailing list which works
in a way of tcp-implementors, but on the topic of socket API
extensions for Infiniband, then I am interested.
-- Pete
On Wed, 20 Feb 2002, Pete Zaitcev wrote:
> > Interested people should contact the working group co-chairs for details
> > about how to join the effort:
> > David Edmondson ([email protected]) or
> > Satya Sharma ([email protected])
> >
> > The URL of the OpenGroup Interconnect Software Consortium is:
> > http://www.opengroup.org/icsc/
>
> Yea, yea. The little problem here is that TOG is not getting it.
> When I see a website that says "Employees of member companies should
> send mail to the Consortium Administrator requesting access to the
> Member area.", I sure as hell is not getting a hardon about working
> on "overtaking Winsock" (and why exactly would we want to copy
> broken attempts to embrace and extend? Perhaps you should talk to
> a certain <i>monkey lover company</i> about it).
Is that, by any chance, the same crowd that stands behind DAFS?
If it _is_ the same crowd - send them to hell, they are beyond hope.
> I would not mind to participate in something more open.
> If anyone has a suggestion of a mailing list which works
> in a way of tcp-implementors, but on the topic of socket API
> extensions for Infiniband, then I am interested.
You need to understand that most of the wacko DMA schemes have already been
laughed out of the IETF - things like TCP RDMA have all turned into
"this way happens to suit my hardware" "we'll we've got a patent on that
way" and other debacles, followed by other people pointing out that they
already get that performance without hacking up protocols and API's.
The existing socket API supports zero copy. SGI proved this a long time back
(Im sure Larry McVoy can give dates). The existing unix aio and real time
signal model supports all the notification needed for efficient scalable
I/O.
You can also implement the entire socket layer in user space on top of
hardware that already does all the brainwork (also been done).
In fact I have a submission for what is needed in API changes. Its a blank
piece of paper right now.
I do agree and I ask the proposed chairs to comment on this - that the IETF
should be involved and if there is such a working group it should be an
IETF working group sponsored by the opengroup. That is where all the real
experts are.
Alan
--
"The IETF already has more than enough RFCs that codify the obvious, make
stupidity illegal, support truth, justice, and the IETF way, and generally
demonstrate the author is a brilliant and valuable Contributor to The
Standards Process." -- Vernon Schryver
On Wed, Feb 20, 2002 at 06:21:11PM -0500, Alexander Viro wrote:
> Is that, by any chance, the same crowd that stands behind DAFS?
> If it _is_ the same crowd - send them to hell, they are beyond hope.
I almost thought that was directed at me, as Coda is kind of like a
Distributed version of AFS.
phew, luckily DAFS it isn't related to AFS or Coda ;)
Jan
On Wed, 20 Feb 2002, Jan Harkes wrote:
> On Wed, Feb 20, 2002 at 06:21:11PM -0500, Alexander Viro wrote:
> > Is that, by any chance, the same crowd that stands behind DAFS?
> > If it _is_ the same crowd - send them to hell, they are beyond hope.
>
> I almost thought that was directed at me, as Coda is kind of like a
> Distributed version of AFS.
>
> phew, luckily DAFS it isn't related to AFS or Coda ;)
DAFS isn't a filesystem - it's software equivalent of Freddy Kruger:
85 madmen^Wcompanies, one naive nurse^Widea, nightmares-inflicting monster
conceived as the result of weeks of clusterfuck...
On Wed, Feb 20, 2002 at 07:00:08PM -0500, Alexander Viro wrote:
>
>
> On Wed, 20 Feb 2002, Jan Harkes wrote:
>
> > On Wed, Feb 20, 2002 at 06:21:11PM -0500, Alexander Viro wrote:
> > > Is that, by any chance, the same crowd that stands behind DAFS?
> > > If it _is_ the same crowd - send them to hell, they are beyond hope.
> >
> > I almost thought that was directed at me, as Coda is kind of like a
> > Distributed version of AFS.
> >
> > phew, luckily DAFS it isn't related to AFS or Coda ;)
>
> DAFS isn't a filesystem - it's software equivalent of Freddy Kruger:
> 85 madmen^Wcompanies, one naive nurse^Widea, nightmares-inflicting monster
> conceived as the result of weeks of clusterfuck...
Yeah. No kidding. Anyone who thinks this is a good idea really hasn't
thought about it clearly. It's nice marketing, sounds good, and is
a crock. If you want this sort of thing, that's what kiobufs are for,
and if kiobufs don't work right, that's a bug. Phooey on hacks.
--
---
Larry McVoy lm at bitmover.com http://www.bitmover.com/lm
> The existing socket API supports zero copy. SGI proved this a long time back
> (Im sure Larry McVoy can give dates).
Vernon was there long before I was and it worked long before I got there,
all I did was hook up the networking plumbing to the file system plumbing
so that you could move data without copying it.
--
---
Larry McVoy lm at bitmover.com http://www.bitmover.com/lm
Alan Cox <[email protected]> writes:
> In fact I have a submission for what is needed in API changes. Its a blank
> piece of paper right now.
Standardizing the linux asynchronous socket error notifications (MSG_ERRQUEUE,
sock_extended_err) would be nice. I guess they could be useful for other
OS too.
-Andi