A Sensible Approach to Speculative Automatic Parallelization

Abstract

The promise of automatic parallelization, freeing programmers from the error-prone and time-consuming process of making efficient use of parallel processing resources, remains unrealized. For decades, the imprecision of memory analysis limited the applicability of automatic parallelization. The introduction of speculation to automatic parallelization overcame these applicability limitations but caused profitability problems due to high communication and bookkeeping costs for speculation validation and commit.

This dissertation shifts the focus from applicability to profitability by making speculative automatic parallelization more efficient. Unlike current approaches that perform analysis and transformations independently and in sequence, the proposed system integrates, without loss of modularity, memory analysis, speculative techniques, and enabling transformations.

Specifically, this dissertation involves three main contributions. First, it introduces a novel speculation-aware collaborative analysis framework (SCAF) that minimizes the need for high-overhead speculation. Second, it proposes a new parallelizing-compiler design that enables careful planning. Third, it presents new efficient speculative privatization transformations that avoid privatization overheads of prior work.

SCAF learns of available speculative information via profiling, computes its full impact on memory dependence analysis, and makes this resulting information available for all code optimizations. SCAF is modular (adding new analysis modules is easy) and collaborative (modules cooperate to produce a result more precise than the confluence of all individual results). Relative to the best prior speculation-aware dependence analysis technique, SCAF dramatically reduces the need for expensive-to-validate memory speculation in the hot loops of all 16 evaluated C/C++ SPEC benchmarks.

The introduction of planning in the compiler enables the selection of parallelization-enabling transformations based on their cost and overall impacts. The benefit is twofold. First, the compiler only applies a minimal-cost set of enabling transformations. Second, this decision-making process makes the addition of new efficient speculative enablers possible, including the proposed privatization transformations.

This dissertation integrates these contributions into a fully-automatic speculative-DOALL parallelization framework for commodity hardware. By reducing speculative parallelization overheads in ways not possible with prior parallelization systems, this work obtains higher overall program speedup (23.0x for 12 general-purpose C/C++ programs running on a 28-core shared-memory machine) than Privateer (11.5x), the prior automatic speculative-DOALL system with the highest applicability.