Rainbow-II Metropolitan-Area Network
This is a joint project between IBM and the Los Alamos National
Laboratory. IBM will be developing the network technology
to be deployed in an applications testbed at Los Alamos.
Rainbow-II is a metropolitan-area network (MAN) supporting upto
32 nodes, at 1 Gb/s per node. Each node is realized in the
form of an optical network adaptor box (ONA) that provides connectivity
to hosts via the HIPPI (high-performance parallel interface) on one
side and the optical network on the other side.
This document is organized into four parts:
The optical network architecture itself is the same as the
Rainbow-I
architecture, shown in this
postscript figure
and described here briefly.
Each station has an
optical transmitter that emits light at a wavelength different
from other transmitters in the network. A fiber leads from each station
to the network hub, which is an all-glass, totally
passive star coupler with N
inputs and N outputs, N being the number of stations in
the network. The star coupler combines the transmissions from different
stations, and at each output, we obtain approximately (1/N)th
of the optical power from each transmitter. From the star coupler, a fiber
leads to each station. Thus, the network is broadcast in nature; all
transmissions reach all the stations. At each station, a tunable
optical receiver selects one of the N wavelengths.
The
optical network adaptor
performs three functions:
- It provides an interface to hosts via HIPPI, the 800 Mb/s
high-performance parallel interface.
- It provides an interface to the
Rainbow
network.
- It provides a number of hardware assists to enable the entire
protocol stack up to the transport layer to be implemented onboard.
The goal is to deliver 1 Gb/s throughput to the application.
The protocol stack used in the ONA/Rainbow-II network is shown
here .
Normally the end-to-end transport protocol used is TCP which
cannot deliver gigabit throughput to the application on most machines.
TCP/IP was designed to optimize communication bandwidth not CPU
processing per packet. Our approach is to replace TCP on the host
by a simple host intersocket protocol (SHIP). SHIP provides a socket
appearance to the applications so that applications will only have
to be relinked to a new library.
SHIP is a lightweight protocol that brings the socket calls from
the application down to a protocol offload engine where the calls
are implemented. SHIP is a master-slave protocol (the host
is the master) and performs connection management and flow control
but does not handle error recovery between the host and the offload
engine.
In this case the protcol offload engine is the ONA,
which runs its own transport protocol called OTP (optical transport
protocol) between ONA boxes over the Rainbow network.
OTP is also a lightweight protocol designed with easy-to-parse headers,
and hardware calculated checksums.
The demonstration at SC'94 features a network of 3 ONAs with RS6000
host machines with attached HIPPI cards
(see figure)
The network is used to set up two concurrent sessions.
One session runs an MPEG video application on top of the SHIP protocol
stack.
The other session transfers high-resolution images across the network.
This session uses a passthrough mode of the ONAs wherein no protocol
offloading is done on the ONAs and the HIPPI packets arriving at
one ONA are simply forwarded over to the other ONA.