Nicolas Arango
MIT EECS - Lincoln Laboratory Undergraduate Research and Innovation Scholar
Task Dependence Prediction for Irregular Parallel Programs
2014–2015
Daniel Sanchez
While enabling the parallelization of programs with consistently independent operations, modern architectures fail to exploit parallelism in irregular computation. To exploit irregular parallelism, hardware and runtime are designed together enabling a software-exposed, task-based speculation scheme with an interface enabling hardware to shoulder the burden of speculative execution of short tasks. The objective of this SuperUROP is the design of the task dependence predictor. To effectively speculate on task execution it is necessary to predict dependencies between tasks and avoid the simultaneous execution of a task and its dependency. Accurate predictions of task dependencies are critical to efficient operation of the proposed architecture.
Computation Structures (6.004) and Computer System Archetecture (6.823) classes piqued my interest in computer architecture and alternatives to conventional computation architectures. A sumer internship designing a DSP and FPGA co-processing system for audio applications opened my eyes to the possibilities of new approaches, fundamentally different than conventional CPUs, for efficient computation.