Parallel packet switching by packet distribution
Problem
Line speeds continue to increase faster than do memory and logic speeds. With
future 40 or 100-Gbps Ethernet, the line speed will begin to exceed memory
bandwidth even for an input buffered switch architecture. One solution
is a parallel packet switch (PPS) with internal switch planes running at a
rate less than line speed. The key problem in a PPS is how packets should be
distributed between planes so that packet ordering is maintained.
Objective
Parallel Packet Switches (PPS) use internal, parallel switch planes that
operate at less than line speed. A PPS can scale-up to faster line speeds than
a single-plane switch can. Load balancing between planes and providing QoS to
flows are open problems. Our objective is to investigate and evaluate methods
of implementing fast and scalable per-packet distribution in a PPS.
Work completed
We have proposed and evaluated the performance of a new "VIQ PPS" that
distributes cells (segmented packets) to multiple planes and uses multiplexors
with novel virtual input queues (VIQ). A VIQ at an output multiplexor consists
of one FIFO queue for each input. It was found that at high offered loads the
mean delay of a VIQ PPS switch is lower than that of a single-plane iSLIP switch
and of all other existing feasible-to-implement PPS designs.
Work in progress
We are currently implementing in FPGA technology the mutliplexor and
demultiplexor of the VIQ PPS. We are also exploring phased-memory designs
to enable low-speed DRAM to be used with high-speed links (i.e., where the
link speed exceeds the memory read and write speed). We are applying for
a small grant from Cisco's
University Research
Program to continue this work beyond the scope of this CAREER project.
Results
This material is based upon work supported by the National Science
Foundation under grant No. 9875177. Any opinions, findings and conclusions or
recommendations expressed in this material are those of the author and do not
necessarily reflects the views of the National Science Foundation (NSF).
Last updated by
Ken Christensen on DECEMBER 20, 2003