Speaker: Evgenij Belikov from Heriot-Watt University

Title: Managing Advisory Parallelism in a Distributed Graph Reducer using Spark
Colocation

Abstract: Work stealing is an adaptive decentralised load balancing mechanism
used in many parallel language run-time systems. A key work stealing decision is
the choice of the spark, which represents potential parallelism, to donate in
response to a received request for work. Commonly, the oldest spark is donated
as it often corresponds to a sub-computation with relatively large granularity.
This talk explores the effect of colocating sparks based on the maximum prefix
matching on encodings of their position within the computation, rather than
using their age, on execution time, stealing success, and fragmentation of the
virtual shared heap. A comparison is made to the default mechanism used in the
GUM run-time system for Glasgow parallel Haskell.