[ut.ee] A Load Balancing Facility :cider Friday 19 - Peirre Delisle

dunc@eecg.toronto.edu (Duncan Elliott) (01/13/90)

                Electrical Engineering Computer Group
                         Cider Seminar Series

          A Load Balancing Facility for Distributed Systems

                            Pierre Delisle
                           Computer Science
                        University of Toronto

         Time: Friday, Jan. 19, 1990, 12:05 --- Place: GB 220

       The advent of distributed systems has created new challenges in
  the  area  of  resource  sharing.  While  the  sharing  of  data and
  peripherals is a fairly common occurrence,  intelligent  sharing  of
  processing  resources  is  not.  Yet, experience has shown that at a
  given instant in time, some nodes sit virtually  idle  while  others
  are  heavily  loaded.  Also,  the  heterogeneity  often found in the
  hardware  configuration  of   distributed   systems   presents   the
  opportunity  for  more effective sharing of the system workload. For
  these reasons, transparent and efficient access  to  the  processing
  resources  of  other  machines  has become increasingly desirable in
  order to improve overall system performance.

  In this talk, we present the design, implementation  and  evaluation
  under  live  workloads  of a general purpose load balancing facility
  for loosely-coupled distributed systems. The load balancing facility
  works  at  the  shell  level,  and  requires  no modification to the
  operating  system  or  application  software.  The  load   balancing
  algorithm  employs  a multi-class eligibility scheme (where resource
  requirements of eligible jobs may be specified) and a multi-resource
  information   scheme  (where  multiple  load  indices  are  used  to
  represent contention at each node) to improve  placement  decisions.
  The  flexibility  of  the implementation makes it relatively easy to
  take into account the heterogeneity of the network, and to adapt  it
  to a wide variety of distributed environments.

  Performance data obtained under live workloads clearly indicate that
  dynamic load balancing at the job level can be easily implemented as
  a system facility in loosely-coupled distributed systems,  and  that
  it  has  the  potential  to  substantially  improve  overall  system