Re: [Nbd] NBD using SDP
- To: "Wouter Verhelst" <w@...112...>
- Cc: email@example.com
- Subject: Re: [Nbd] NBD using SDP
- From: "Mike Snitzer" <snitzer@...17...>
- Date: Tue, 16 Oct 2007 08:44:39 -0400
- Message-id: <170fa0d20710160544n2e18c56cmb0a78c05df356a3d@...18...>
- In-reply-to: <20071016105002.GA4218@...39...>
- References: <1192490510.4558.2.camel@...140...> <20071016105002.GA4218@...39...>
On 10/16/07, Wouter Verhelst <w@...112...> wrote:
> On Mon, Oct 15, 2007 at 06:21:50PM -0500, John Leidel wrote:
> > All... apologies for the random post...
> It's not random ;-)
> > sometime ago I created a simple patch set for NBD to directly utilize
> > the RDMA capabilities of Infiniband via the Sockets Direct Protocol.
> What are the benefits of doing so? I guess I could be pursuaded to
> accept this into nbd-tools mainline, but I'd really have to know what
> these protocols do, and how they benefit NBD.
NBD would be more performant on an Infiniband network because SDP
allows more efficient use of the IB stack than if relying on IPoIB.
The cpu overhead should be reduced and overall throughput increased.
BTW, if you were to accept this wouldn't you just be patching the nbd
source directly (as is indicated by John's tarball of patches)?
Adding SDP support implies altering the socket address family to be
AF_INET_OFFLOAD rather than AF_INIT. So SDP support could easily be
made a configure/compile time thing.