Drew Eckhardt <[EMAIL PROTECTED]> wrote:
> IP over SCSI is problematic because of per-command overhead and the
> often small packet sizes.
Right. However, what prevents the ``network device'' from clustering
packets ? Something like:
- we always send everything we have (no delay: ie the first I/O is
of 1.5 kb or something, only)
- but we accept any packet even when transmitting (upto a certain
queue length). To avoid skb congestion (too many packets to be
sent), data could be copied. I don't know if this is useful.
So the next I/O is much bigger. This of course needs some changes
in the drivers.
This would allow a very big speed, without too much overhead, and
without needing to have bigger packet sizes (so no fragmentation
needed if SCSI -> Ethernet, for example).
> For certain problems (Beowulf clusters), cable length isn't a real
> issue.
And modern SCSI is around 13 m length. Which is either very long (for
a cluster), or very short (for a building).
-
To unsubscribe from this list: send the line "unsubscribe linux-scsi" in
the body of a message to [EMAIL PROTECTED]
