Extending the Nested Parallel Model to the Nested Dataflow Model with Provably Efficient Schedulers

TL;DR

This paper introduces a new dataflow construct to extend the nested parallel model, enabling more expressive algorithms with optimal performance and cache efficiency on multicore architectures, supported by provably efficient schedulers.

Contribution

It proposes the nested dataflow model with a new composition construct, designs optimal algorithms within this model, and develops schedulers that guarantee locality and load balancing.

Findings

Algorithms in the ND model achieve optimal span and cache complexity.

SB schedulers can utilize increased parallelism for better cache and time bounds.

The running time scales with the sum of cache complexities divided by processors.

Abstract

The nested parallel (a.k.a. fork-join) model is widely used for writing parallel programs. However, the two composition constructs, i.e. "$\parallel$" (parallel) and "$;$" (serial), are insufficient in expressing "partial dependencies" or "partial parallelism" in a program. We propose a new dataflow composition construct "$\leadsto$" to express partial dependencies in algorithms in a processor- and cache-oblivious way, thus extending the Nested Parallel (NP) model to the \emph{Nested Dataflow} (ND) model. We redesign several divide-and-conquer algorithms ranging from dense linear algebra to dynamic-programming in the ND model and prove that they all have optimal span while retaining optimal cache complexity. We propose the design of runtime schedulers that map ND programs to multicore processors with multiple levels of possibly shared caches (i.e, Parallel Memory Hierarchies) and provide theoretical guarantees on their ability to preserve locality and load balance. For this, we adapt space-bounded (SB) schedulers for the ND model. We show that our algorithms have increased "parallelizability" in the ND model, and that SB schedulers can use the extra parallelizability to achieve asymptotically optimal bounds on cache misses and running time on a greater number of processors than in the NP model. The running time for the algorithms in this paper is $O\left(\frac{\sum_{i=0}^{h-1} Q^{*}({\mathsf t};\sigma\cdot M_i)\cdot C_i}{p}\right)$, where $Q^{*}$ is the cache complexity of task ${\mathsf t}$, $C_i$ is the cost of cache miss at level-$i$ cache which is of size $M_i$, $\sigma\in(0,1)$ is a constant, and $p$ is the number of processors in an $h$-level cache hierarchy.