Re: Requirements specification

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

> I don't think network channel speeds are increasing briskly. Ethernet was 10
> Mb/s in 1980, 100 Mb/s in 1990, and 1 Gb/s in 2000. These are slow
> infrastructural step functions. 

Past performance is not necessarily indicative of future results.

The reason the generation cycle for edge channel technology (Ethernet)
has not been faster over the past 20 years is lack of demand.

In the 80s and 90s the giants were platform vendors, and the bulk of
the smaller hardware companies also produced platform-related things.
Not many of those companies left.  On the other hand, there are
zillions of I/O and networking companies now.  Clearly, this is
because the demand for networking has increased dramatically.  This
will inevitably increase the rate at which the technical hurdles to
faster channel speeds are overcome.

Pipe size is presently one of the least of these hurdles.  T11.1 has
seen various feasible proposals for physical layers which deliver a
substantial fraction of typical platform memory bandwidth (which is
the ultimate `speed limit'), for increasingly fast platforms. 

The most immediate hurdle is what iSCSI is addressing---efficient data
delivery at faster speeds.  There's no point in having a fast channel
if you can't make use of it.  Once appropriate hardware acceleration
is available, the demand for faster channels will come as applications
are developed to use it, and the physical layer technologies are close
at hand.

> In the 1990's there was a transition from shared bandwidth to switched
> bandwidth which helped mitigate the problem of the slow evolution of link
> speeds.

The biggest problem this began to solve was how to introduce faster
channels into the network infrastructure.  If you have a shared
medium, it is very difficult to support channels of different speeds.
This is one of the reasons why faster FC is slow in coming, even
though the physical medium technology has been around.  If the medium
is not shared, you can upgrade on a link by link basis.

The other hurdle that has to be crossed once you allow channels of
various speeds is congestion.  Any infrastructure that lacks some form
of congestion control (adaptive (TCP), or bandwidth reservation (ATM))
will not be able to support increasing link speeds effectively.  This
is another reason why faster FC is slow in coming.

What iSCSI brings to the table (assuming it can carry it without
dropping it :^), is hardware data handling with congestion control and
media independence, hopefully in a way that other protocols besides
storage can also use the hardware.  This is the dynamite which will
blast open the door to memory speed channels.  Once you have a memory
speed channel, there's no need to stripe.

Steph