I have a feeling that this will become a popular idea soon largely due
flexibility provided by the efficient routing implementation. I've had
similar ideas / feelings about using lnet apart from lustre. One of the
obvious drawbacks is that the interesting LNDs require kernel mode
execution to use them (I think this is still true??). As far as API
changes, I'd suspect that lustre itself would be very sensitive to LNET
changes. If user's of LNET are no less sensitive than lustre itself
then I think it could be useful to people.
paul
Eric Barton wrote:
Dave,
I enjoyed chatting with you too - exciting stuff!
I'm afraid I'm completely schizophrenic about pushing LNET as a
general purpose communications API.
My right brain says LNET is an _ideal_ networking platform for all
sorts of high performance distributed applications running on
inhomogeneous networks with various support for RDMA.
But my left brain reminds me that making LNET a supported standard is
kind of at odds with the reason we forked it from Sandia
Portals, which was to allow its API to change according to lustre's
needs for a networking abstraction. The introduction of further RAS,
usability and scalability features into lustre must inevitably affect
all layers of the stack which would make it a moving target for
independent use.
I wonder if people on lustre-discuss has a view.
Cheers,
Eric
------------------------------------------------------------------------
*From:* Cohen, David (GTI) [mailto:[EMAIL PROTECTED]
*Sent:* 18 November 2006 7:01 PM
*To:* [EMAIL PROTECTED]
*Cc:* [EMAIL PROTECTED]; [EMAIL PROTECTED];
[EMAIL PROTECTED]; [EMAIL PROTECTED];
[EMAIL PROTECTED]
*Subject:* portals/lnet as an abstraction over rdma and/or verbs
Hi Eric,
It was great meeting you down in Tampa this past week. I really
enjoyed your presentation on lnet as well as talking about some of
the interesting challanges we face with our electronic trading
business.
Am contacting you here as I am curious on the applicability of
using lnet independently of Lustre, basically as a general purpose
messaging abstraction over rdma and verbs. This could, for
example, then be wrapped by something like ATT's sfio or be
surfaced from within Java as an NIO transport plug-in.
Such an approach would allow for leverage of the abstraction lnet
provides over independent transport mediums like myrinet,
infiniband, 10 gbe, etc while surfacing a reasonably simple
channel/stream abstraction to higher level applications. This
would eliminate the need for something like SDP, as an example.
Thanks in advance for your thoughts on this,
/dave.
------------------------------------------------------------------------
If you are not an intended recipient of this e-mail, please notify
the sender, delete it and do not read, act upon, print, disclose,
copy, retain or redistribute it. Click here
<http://www.ml.com/email_terms/>for important additional terms
relating to this e-mail. http://www.ml.com/email_terms/
------------------------------------------------------------------------
------------------------------------------------------------------------
_______________________________________________
Lustre-discuss mailing list
[email protected]
https://mail.clusterfs.com/mailman/listinfo/lustre-discuss
_______________________________________________
Lustre-discuss mailing list
[email protected]
https://mail.clusterfs.com/mailman/listinfo/lustre-discuss
