Re: TCP limitations (was Re: ISCSI: Urgent Flag requirement violates TCP.)

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

Vern,

Thanks for articulating this property.

> So if you want to go fast, you have to have *much less* than one sequence
> hole per 10^4 packets.

This is only assuming you want to go fast on a single TCP stream.
Certainly 1.8 MB/s seems disappointing for a single stream, but 100 MS
RTT will cover a lot of ground, too.

In this WAN case, one realistic scenario is shared access to some sort
of substantial storage farm, which will mean many streams.

As you have mentioned, this isn't an iSCSI problem, per se.  Any bulk
data access over long latency connections with TCP has this problem,
and iSCSI should be (is) positioned to leverage whatever solutions are
developed.  For example, I suppose Akamai's claim that the solution is
to have the data close to the request applies fully to iSCSI (don't
cross the beams).  So would enhancements to TCP, or use of another
transport layer which better addresses the problem.  Not too long ago,
TCP was theoretically incapable of delivering gigabit performance at
all.  The transport layers don't stand still.  

> But the Urgent pointer only helps when you happen to have a sequence
> hole.  So how in practice can it be worth the effort?

It will permit a high speed endpoint to deliver full pipe aggregate
throughput across multiple streams.

Steph