The architects of parallel and distributed systems like PC clusters can choose among approximately half a dozen fairly established Gigabit/s interconnect technologies. Most of them are shipped as a add-on cards that plug directly into the PCI bus of the PC motherboards used as nodes in such systems.
In this talk I am presenting two recent research contributions in the area of Gigabit Networking: First I will discuss a model that serves
as a common ground to compare the different interconnect architectures in several different settings between coarse grain and fine grain
communication. I will take a closer look at the architecture and the
performance characteristics of the two examples: Myrinet and Dolphin SCI and compare them to the interconnect of a Cray T3D, an old but
very fast MPP. Second I will introduce "speculative defragmentation", a highly sophisticated software architecture to speed up
GigabitEthernet network processing on machines that are slower than the network requiring a true zero copy protocol architecture in order
to deliver a reasonable fraction of a Gigabit in throughput.
The software tricks around speculative fragmentation will lead to a critical review of Ethernet as a high speed interconnect and permit a conclusion with a outlook to future interconnects like Infiniband.
Projects: www.cs.inf.ethz.ch/CoPs
Slides: www.cs.inf.ethz.ch/CoPs/speedup