Infinipath

Introduction
HPC Architecture
  1. Shared-memory SIMD machines
  2. Distributed-memory SIMD machines
  3. Shared-memory MIMD machines
  4. Distributed-memory MIMD machines
  5. ccNUMA machines
  6. Clusters
  7. Processors
    1. AMD Magny-Cours
    2. IBM POWER6
    3. IBM POWER7
    4. IBM PowerPC 970MP
    5. IBM BlueGene processors
    6. Intel Xeon
    7. The SPARC processors
  8. Accelerators
    1. GPU accelerators
      1. ATI/AMD
      2. nVIDIA
    2. General accelerators
      1. The IBM/Sony/Toshiba Cell processor
      2. ClearSpeed/Petapath
    3. FPGA accelerators
      1. Convey
      2. Kuberre
      3. SRC
  9. Networks
    1. Infiniband
    2. InfiniPath
    3. Myrinet
Available systems
  • The Bull bullx system
  • The Cray XE6
  • The Cray XMT
  • The Cray XT5h
  • The Fujitsu FX1
  • The Hitachi SR16000
  • The IBM BlueGene/L&P
  • The IBM eServer p575
  • The IBM System Cluster 1350
  • The NEC SX-9
  • The SGI Altix UV series
    		•
    Systems disappeared from the list
    Systems under development
    Glossary
    Acknowledgments
    References

    InfiniPath only provides Host Channel Adapters with a 4-wide (1.25 GB/s) Infiniband link on the network side and connecting to a HyperTransport bus or PCI-Express at the computer side. For systems with AMD processors on board the HyperTransport option is particularly attractive because of the direct connection to the host's processors. This results in very low latencies for small messages. PathScale, the vendor of the InfiniPath HCAs quotes latencies as low as 1.29 µs. Obviously, this type of HCA cannot be used with systems based on non-AMD processors. For these systems the HCAs with PCI-Express can be used. They have slightly higher, but still low latency of 1.6 µs. The effective bandwidth is also high: a uni-directional bandwidth of ≅ 950 MB/s can be obtained using MPI for both types of HCA.

    The InifiniPath HBAs do not contain processing power themselves. Any processing associated with the communication is done by the host processor. According to PathScale this is an advantage because the host processor is usually much faster than the processors employed in switches. An evaluation report from Sandia National Lab [8] seems to corroborate this assertion.
    PathScale only offers HCAs (and the software stack coming with it) and these can be used by any Infiniband switch vendor that adheres to the OpenIB protocol standard which are pretty much all of them.