Reliable Scalable Symbolic Computation: The Design of SymGridPar2

Phil Trinder


Symbolic computations are challenging to parallelise as they have
complex data and control structures, and both dynamic and highly
irregular parallelism.  The SymGridPar framework has been developed to
address these challenges on small-scale parallel architectures.
However, as the number of cores in compute clusters continues to grow
exponentially, and as the communication topology is becoming
increasingly complex, an improved parallel symbolic computation
framework is required: SymGridPar2.

In this talk, I'll explain how two main aspects of the design of
SymGridPar2, fault tolerance and locality control, interact with
dynamic scheduling of parallelism.  Fault tolerance is achieved by
tracking the location of tasks as they are scheduled across the
network, and by replicating tasks that were affected by node failure.
Locality control exposes an abstraction of the communication topology
so programs can control how close together tasks shall be placed by
the dynamic scheduler.