Re: Unsolicited Data.

[Stephen Bailey <steph@cs.uchicago.edu>]

Matthew,

> However, for large writes the unsolicited portion of the data need
> not be all the data in the transfer but enough to overcome the
> initial latency.

I was probably unclear (well, OK, completely incoherent) in my
previous message.  Lemmie try it again.

My point about large writes is that there are usually many of them
outstanding concurrently.  That's what file systems live for.  The
initial latency of the next write is covered by the terminal data
transfer of the current write.

As long as there's bandwidth * delay worth of total outstanding
writes, you don't need unsolicited data.  If you have an obscene
bandwidth * delay, you might realistically expect that current file
systems WON'T give you that much outstanding demand, so saving the
startup latency with unsolicited data will improve performance
somewhat.  But not much.  I'm claiming that's not a case worth
optimizing.

Steph