Date: Fri, 14 Apr 1995 21:55:41 -0700 From: "Russell L. Carter" <rcarter@geli.com> To: phk@ref.tfs.com, rgrimes@gndrsh.aac.dev.com Cc: freebsd-hackers@FreeBSD.org, julian@TFS.COM Subject: Re: Just how fast can we go... (was: Re: SCSI target) Message-ID: <199504150455.VAA02893@geli.clusternet>
next in thread | raw e-mail | index | archive | help
|> [cc: trimmed to hackers, seems the best place, since we are talking |> about both scsi and 100MB/sec ether :-)] |> > |> > > |> > > > |> > > Well, with 100MB/S ethernet support now being a reality TCP/IP over |> > > SCSI only has an advantage for *wide* scsi controllers. |> > weeeellll, no, there are advandages in being able to transfer |> > 128KB of scatter-gather data in one hit with NO |> > cpu intervention..... :) |> |> ... |> |> The 21040/21140 chips are bus master just like a scsi controller, and |> can infact do some very large scatter-gather's in there own right. | |But you are still limited to the packet-size on the 100mb net... | If you are going to do a "nice little cluster", there is nothing more important than latency. Bandwidth doesn't matter so much, once you get above several user megabytes per second. I am not able to judge how gather-scatter affects things; I do know that nearly all modern numerical algorithms are constructed so that this does not occur. CRI systems for quite some time have had hardware support for gather-scatter and I've never seen an intelligently written code that required it. YMMV. Russell
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?199504150455.VAA02893>