An Adaptive Framework for Large-Scale State Space Search

Abstract

Abstract-State space search problems abound in the artificial intelligence, planning and optimization literature. Solving such problems is generally NP-hard. Therefore, a brute-force approach to state space search must be employed. It is instructive to solve them on large parallel machines with significant computational power. However, writing efficient and scalable parallel programs has traditionally been a challenging undertaking. In this paper, we analyze several performance characteristics common to all parallel state space search applications. In particular, we focus on the issues of grain size, the prioritized execution of tasks and the balancing of load among processors in the system. We demonstrate the techniques that are used to scale such applications to large scale. We have incorporated these techniques into a general search engine framework that is designed to solve a broad class of state space search problems. We demonstrate the efficiency and scalability of our design using three example applications, and present scaling results up to 16,384 processors.

On scalable parallel recursive backtracking

“…Several models have already been presented in the literature including centralized (i.e. the master-worker(s) model where most of the communication and task distribution duties are assigned to a single core) [18], decentralized [15,17], or a hybrid of both [10]. Although each model has its pros and cons, centralization rapidly becomes a bottleneck when the number of computing cores exceeds a certain threshold [18].…”

“…Note that this indexing method can easily be extended for arbitrary branching factor by simply setting the index of the k th child of N d,p to idx(N d,p ) + "(k − 1)". The general idea of indexing is not new and has been previously used for prioritizing tasks in buffers or queues [10,13]. However, as we shall see next, we can completely eliminate the need for buffering multiple tasks by combining a fully decentralized communication model with some operations for manipulating indices.…”

“…Then, each core C i is assigned subtree T i . This seemingly straightforward parallel nature of search tree decomposition is deceiving: previous work has shown that attaining scalability is far from easy [9,10,11,12].…”

Optimizing Data Locality for Fork/Join Programs Using Constrained Work Stealing

“…Static compile-time approaches [31] have been employed. Charm++ has adaptive grain size control for state space search [32]. Other work has focused on only splitting grain sizes when a worker is in need of work, which is feasible with a managed runtime [33], [34], [35].…”

A distributed dynamic load balancer for iterative applications

“…There is another class of applications, such as combinatorial search, that involves dynamic creation of work and therefore has a tendency for imbalance. This class of applications has distinct characteristics and load balancing needs, and has been addressed by much past work such as work-stealing [25,3,32]. This paper does not focus on such applications but instead on the iterative applications, which are predominant in science and engineering.…”