Commodity OS Bypass and Flow Control Issues

April 24th, 1998

Author: Ian Philp

Commodity OS Bypass and Flow Control Issues

OS Bypass

OS Bypass and Flow Control

Flow Control and ULP Issues

VIA and ST APIs


Slide 1 – Commodity OS Bypass and Flow Control Issues

philp01
  • Network Engineering Group
  • Los Alamos National Laboratory
  • SOS DVS Workshop
  • March 30 – April 1, 1998

Slide 2 – OS Bypass

philp02
  • latency
    • avoid system calls and interrupts
    • requires NIC to mux/demux application messages
  • bandwidth
    • avoid memory-memory copies
  • VIA and ST are both OS bypass solutions
    • provide interrupt and polling interface
    • coalesce interrupts

Slide 3 – OS Bypass and Flow Control

philp03
  • tradeoff between
    • price of NIC
    • potential performance improvement
  • flow control on the NIC
    • decouples progress of messages from application scheduling on CPU

Slide 4 – Flow Control and ULP Issues

philp04
  • Sender
  • Receiver
  • MPI_ISend()
  • compute()
  • MPI_IRECV()
  • Data with NIC
  • flow control (ST)
  • Data with appl.
  • flow control (VIA)
  • NIC
  • request-to-send
  • clear-to-send

Slide 5 – VIA and ST APIs

philp05
  • VIA
  • ST
  • MPI
  • sockets
  • MPI
  • sockets
  • VIA
  • VIA
  • application
  • application
  • ST
  • host
  • host
  • NIC
  • NIC