Hierarchical memory management for parallel programs

Abstract

An important feature of functional programs is that they are parallel by default. Implementing an efficient parallel functional language, however, is a major challenge, in part because the high rate of allocation and freeing associated with functional programs requires an efficient and scalable memory manager. In this paper, we present a technique for parallel memory management for strict functional languages with nested parallelism. At the highest level of abstraction, the approach consists of a technique to organize memory as a hierarchy of heaps, and an algorithm for performing automatic memory reclamation by taking advantage of a disentanglement property of parallel functional programs. More specifically, the idea is to assign to each parallel task its own heap in memory and organize the heaps in a hierarchy/tree that mirrors the hierarchy of tasks. We present a nested-parallel calculus that specifies hierarchical heaps and prove in this calculus a disentanglement property, which prohibits a task from accessing objects allocated by another task that might execute in parallel. Leveraging the disentanglement property, we present a garbage collection technique that can operate on any subtree in the memory hierarchy concurrently as other tasks (and/or other collections) proceed in parallel. We prove the safety of this collector by formalizing it in the context of our parallel calculus. In addition, we describe how the proposed techniques can be implemented on modern shared-memory machines and present a prototype implementation as an extension to MLton, a highperformance compiler for the Standard ML language. Finally, we evaluate the performance of this implementation on a number of parallel benchmarks.

Competitive parallelism: getting your priorities right

“…We have developed a prototype implementation of PriML. Our implementation compiles PriML to mlton-parmem (Raghunathan et al, 2016), a parallel extension of Standard ML which is derived from the work of Spoonhower (2009). We have also developed a parallel scheduler for PriML programs, which plugs into the mlton-parmem runtime.…”

“…Cooperatively threaded languages such as NESL (Blelloch et al, 1994), Cilk (Frigo et al, 1998), parallel Haskell (Chakravarty et al, 2007Keller et al, 2010) and parallel ML (Fluet et al, 2011;Jagannathan et al, 2010;Raghunathan et al, 2016), have at least two important features:…”

Hierarchical memory management for mutable state

“…As in Raghunathan et al [31], to support parallel automatic memory management, each task is assigned its own heap in which it allocates new data. Heaps are organized into a hierarchy (tree) with the same parent/child relationships as their associated tasks.…”

“…The execution of msort in Figure 2 is disentangled, thus the two leaves of the hierarchy could both be independently collected in parallel; for example, collecting the array a after computing Seq.fromArray a but before joining with the parent task. It can be shown that all purely functional programs naturally exhibit disentanglement [31].…”

“…In this paper, we propose techniques for supporting mutable data efficiently within the run-time system of nestedparallel functional languages, focusing on strict languages in the style of the ML family. Our approach builds upon that of hierarchical heaps, a memory management technique developed in prior work [31]. The basic idea is to organize memory so that it mirrors the structure of the parallel computation.…”