[e2e] Re:Are you interested in TOEs and related issues

[H.K. Jerry Chu]

>I've trimmed cross-posting a lot in this note.
>
>The thread so far has made numerous assumptions (e.g. small packets
>consume as much buffer space as big packets in real routers) about
>how modern switches/routers are designed.   The validity of many
>of those assumptions is not obvious to me.
>
>So from where I sit, this thread has long since devolved into
>pretty theoretical territory ("if people were silly enough
>to build their routers in a certain way, and the following
>unusual event occurs, then on certain days of the week this
>bad event might happen...").
>
>It might make sense for the folks interested in this thread
>to actually pull back a bit, consider how actual shipping
>switches/routers are built, and restart from there.
>
>Oh, and the subject line probably needs updating, since how one
>builds a switch/router is farther divorced from the questions
>of whether TOEs make technical/economic sense than some folks
>here might believe.[1]

Actually it's very relavent. If switch/router's tolerance on pkt
burst has scaled up like the CPU speed or the line speed (i.e.,
router can handle a burst of 20 pkts just as well as they handled a
burst of 2 pkts 10 years ago), then at least for bulk data, a
stateless offload technology called LSO (large segment offload)
may be suffcient to offload the hosts. You don't need TOEs. LSO
lives and dies on bursts because it doesn't run the TCP algorithm
to know when to properly schedule pkt transmission.

BTW, to many of us in engineering, not marketing TOE refers to a full
TCP offload, including all the algorithms (which is the most
troublesome part). I wouldn't call cksum offload, or even LSO "TOE".

Jerry

>
>Cheers,
>
>Ran
>rja at extremenetworks.com
>
>[1] I don't work on a host/server stack these days, but my past 
>experience
>is pretty closely aligned with Craig's earlier summary of community
>experience with TOEs from the 1980s.
>