[gpfsug-discuss] GPFS Routers

Knister, Aaron S. (GSFC-606.2)[COMPUTER SCIENCE CORP] aaron.s.knister at nasa.gov
Tue Sep 20 15:07:38 BST 2016


Not sure if this image will go through but here's one I found:

[X]

The "Routers" are LNET routers. LNET is just the name of lustre's network stack. The LNET routers "route" the Lustre protocol between disparate network types (quadrics, Ethernet, myrinet, carrier pigeon). Packet loss on carrier pigeon is particularly brutal, though.



From: Marc A Kaplan
Sent: 9/20/16, 10:02 AM
To: gpfsug main discussion list
Subject: Re: [gpfsug-discuss] GPFS Routers
Thanks for spelling out the situation more clearly.  This is beyond my knowledge and expertise.
But perhaps some other participants on this forum will chime in!

I may be missing something, but asking "What is Lustre LNET?" via google does not yield good answers.
It would be helpful to have some graphics (pictures!) of typical, useful configurations.  Limiting myself to a few minutes of searching, I couldn't find any.

I "get" that Lustre users/admin with lots of nodes and several switching fabrics find it useful, but beyond that...

I guess the answer will be "Performance!" -- but the obvious question is: Why not "just" use IP - that is the Internetworking Protocol!
So rather than sweat over LNET, why not improve IP to work better over several IBs?

>From a user/customer point of view where "I needed this yesterday", short of having an "LNET for GPFS", I suggest considering reconfiguring your nodes, switches, storage
to get better performance.  If you need to buy some more hardware, so be it.

--marc



From:        Aaron Knister <aaron.s.knister at nasa.gov>
To:        <gpfsug-discuss at spectrumscale.org>
Date:        09/20/2016 09:23 AM
Subject:        Re: [gpfsug-discuss] GPFS Routers
Sent by:        gpfsug-discuss-bounces at spectrumscale.org
________________________________



Hi Marc,

Currently we serve three disparate infiniband fabrics with three
separate sets of NSD servers all connected via FC to backend storage.

I was exploring the idea of flipping that on its head and having one set
of NSD servers but would like something akin to Lustre LNET routers to
connect each fabric to the back-end NSD servers over IB. I know there's
IB routers out there now but I'm quite drawn to the idea of a GPFS
equivalent of Lustre LNET routers, having used them in the past.

I suppose I could always smush some extra HCAs in the NSD servers and do
it that way but that got really ugly when I started factoring in
omnipath. Something like an LNET router would also be useful for GNR
users who would like to present to both an IB and an OmniPath fabric
over RDMA.

-Aaron

On 9/12/16 10:48 AM, Marc A Kaplan wrote:
> Perhaps if you clearly describe what equipment and connections you have
> in place and what you're trying to accomplish, someone on this board can
> propose a solution.
>
> In principle, it's always possible to insert proxies/routers to "fake"
> any two endpoints into "believing" they are communicating directly.
>
>
>
>
>
> From:        Aaron Knister <aaron.s.knister at nasa.gov>
> To:        <gpfsug-discuss at spectrumscale.org>
> Date:        09/11/2016 08:01 PM
> Subject:        Re: [gpfsug-discuss] GPFS Routers
> Sent by:        gpfsug-discuss-bounces at spectrumscale.org
> ------------------------------------------------------------------------
>
>
>
> After some googling around, I wonder if perhaps what I'm thinking of was
> an I/O forwarding layer that I understood was being developed for x86_64
> type machines rather than some type of GPFS protocol router or proxy.
>
> -Aaron
>
> On 9/11/16 5:02 PM, Knister, Aaron S. (GSFC-606.2)[COMPUTER SCIENCE
> CORP] wrote:
>> Hi Everyone,
>>
>> A while back I seem to recall hearing about a mechanism being developed
>> that would function similarly to Lustre's LNET routers and effectively
>> allow a single set of NSD servers to talk to multiple RDMA fabrics
>> without requiring the NSD servers to have infiniband interfaces on each
>> RDMA fabric. Rather, one would have a set of GPFS gateway nodes on each
>> fabric that would in effect proxy the RDMA requests to the NSD server.
>> Does anyone know what I'm talking about? Just curious if it's still on
>> the roadmap.
>>
>> -Aaron
>>
>>
>> _______________________________________________
>> gpfsug-discuss mailing list
>> gpfsug-discuss at spectrumscale.org
>> http://gpfsug.org/mailman/listinfo/gpfsug-discuss
>>
>
> --
> Aaron Knister
> NASA Center for Climate Simulation (Code 606.2)
> Goddard Space Flight Center
> (301) 286-2776
> _______________________________________________
> gpfsug-discuss mailing list
> gpfsug-discuss at spectrumscale.org
> http://gpfsug.org/mailman/listinfo/gpfsug-discuss
>
>
>
>
>
> _______________________________________________
> gpfsug-discuss mailing list
> gpfsug-discuss at spectrumscale.org
> http://gpfsug.org/mailman/listinfo/gpfsug-discuss
>

--
Aaron Knister
NASA Center for Climate Simulation (Code 606.2)
Goddard Space Flight Center
(301) 286-2776
_______________________________________________
gpfsug-discuss mailing list
gpfsug-discuss at spectrumscale.org
http://gpfsug.org/mailman/listinfo/gpfsug-discuss



-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://gpfsug.org/pipermail/gpfsug-discuss_gpfsug.org/attachments/20160920/52bfbf08/attachment-0002.htm>


More information about the gpfsug-discuss mailing list