Re: Scheduled Transfer Protocol (ST)

[Aman Singla <aman@cthulhu.engr.sgi.com>]

> Thanks for bringing the ST protocol to our attention. I'm not sure
> it fits all the requirements, though, because one of the goals of IPS
> is to generate a WAN-friendly protocol.
> 
> Is STP designed for the WAN? I see no provision in the ST spec for
> TCP-style congestion control or even exponential back-off. In addition,

You are right in pointing out that STP as a protocol is still quite
immature when it comes to applying it to a WAN scenario. What I
would say is that STP has various useful primitives for congestion
control etc. built into it - the number of outstanding CTSs, the
block size to use etc. are all tunable parameter and can be updated
in the STP header by switches/router/entities enroute; there are
primitives to timeout and regenerate requests.
But the mechanisms to trigger timeouts, and determine the various
parameters correctly for a WAN scenario are missing; and require
work/input/design based on the requirements.

> the Clear_To_Send mechanism as currently speced seems to add
> a round-trip time to every transfer, increasing latency.

For significantly sized transfers, the CTSs can be and are pipelined
with the data-transfer; so the only observable CTS latency is for the
first block of data - which ammortized over the length of the transfer
still gives you great b/w.
For latency sensitive short message exchanges, persistently mapped
memory regions act as data buffers and can be incredibly fast by not
requiring CTSs for the data.

> As for the SCSI/ST spec:
> 
> Is there any mechanism for ordering tasks? For example, if you
> want to enter several ordered tasks into a SCSI target, can you
> issue them simultaneously, i.e. without waiting for an ACK?.

I'm not very sure about the point you're trying to make here. STP
will just be the transfer medium for the SCSI commands and the data -
in a way like SCSI fibre channel is..
Yeah! a SCSI stack can issue multiple outstanding requests to a SCSI
target using STP - their ordering requirement, if any, would be
determined by the SCSI header - and their ordering, if any, would
be determined by the SCSI target.
My apologies if I haven't grasped your question correctly..

thanks,

:a