A Comparison of two Gigabit SAN/LAN technologies:
            Scalable Coherent Interface versus Myrinet

                   Ch. Kurmann and T. Stricker
                Laboratory for Computer Systems
              Swiss Institute of Technology (ETH)
                   CH-8092 Zurich, Switzerland
                   kurmann,tomstr@inf.ethz.ch

  Several commercial Gigabit interconnect technologies are currently
  emerging due to the results of the intensive research with the NSF
  Gigabit Networking testbeds. A fair comparison of these new, but
  different network technologies needs a common denominator as a
  basis.  To characterize networking performance, we propose to carry
  out a comparison at three different levels: (1) simple remote
  load/store operations, (2) message passing libraries and (3)
  standard IP over LAN networking.  Our comparison includes the two
  Gigabit Interconnects for the PCI bus that are targeted at clusters
  of PCs, Myrinet and Dophin SCI.  Just for a reference the
  interconnect technology of the SGI/Cray T3D is evaluated with the
  same benchmarks.  After a description of the hardware in three
  different network interfaces, we take a closer look at the
  communication operations and determine the ``native'' and
  ``alternate'' modes of operation of the different technologies.  For
  a direct comparison we present uniform throughput measurements for
  all three architectures at all levels of communication services.
  
  Myrinet and the Cray T3D can sustain a full Gigabit/s transfer rate
  for large transfers of contiguous blocks. The Dolphin SCI cards peak
  at somewhat more than half a Gigabit/s despite the faster raw
  hardware speed. The two PCI based system can not sustain their good
  performance when the communication involves non contiguous data,
  when the message passing semantics require buffering in main memory
  or when an IP over LAN emulation is used. Given the low cost of
  these interconnect boards, their performance close to a Gigabit is
  remarkable, but due to the afore mentioned limitations we expect that
  any Cluster of PCs built with those PCI card interconnects will get
  into some difficulties with applications that require complex, dense
  communication patterns or support for standardized high level
  networkin software.