On data-locality issues of task-based programming models

As the classical technology scaling ended (processor clock rates are not growing anymore), performance increase has to be gained from explicit parallelism. The modern computing is cheap and massively parallel, while energy and performance costs are impacted by data movement: a chip of a flagship supercomputer is expected to have soon thousands cores on it; the energy cost for computation is decreasing much faster than the energy cost for moving data on a chip, thus the latter is becoming a top priority in order to have computing efficiency. Current HPC applications have to be ready for the Exascale era, and if current parallel programming models do not totally support constructs that describe data locality and affinity, but future Exascale programming models have to.

The concept of data locality implies the probability of a memory reference being “local” to prior memory accesses. Among parallel programming models for shared-memory machines, the use of task-based programming models is becoming more and more common. Task-based approach refers to designing a program in terms of “tasks” - a logically discrete section of work to be done.

Here we will talk about different parallel programming models in general, about data-locality issues of task-based programming models, the implications between task granularity and task scheduling overhead on many-core shared-memory systems and others.