Improvement of the Embarrassingly Parallel Search for Data Centers

Abstract

International audienceWe propose an adaptation of the Embarrassingly Parallel Search (EPS) method for data centers. EPS is a simple but efficient method for parallel solving of CSPs. EPS decomposes the problem in many distinct subproblems which are then solved independently by workers. EPS performed well on multi-cores machines (40), but some issues arise when using more cores in a datacenter. Here, we identify the decomposition as the cause of the degradation and propose a parallel decomposition to address this issue. Thanks to it, EPS gives almost linear speedup and outperforms work stealing by orders of magnitude using the Gecode solver

A Comparison of Decomposition Methods for the Maximum Common Subgraph Problem

“…Experiments in [4], [5] show us that a good decomposition is generally obtained, with EPS, by generating about 30 subproblems per worker. In [5], the average speedup reported with EPS is close to 0.5k, where k is the number of workers, on a large benchmark of instances. This is much better than the speedup obtained with a work stealing approach, where subproblems are dynamically generated by splitting the subproblem currently solved by a worker, whenever another worker has finished its own work.…”

“…However, this problem is N P-hard and very challenging. To speed up the solution process, we may decompose the problem into independent subproblems which are solved in parallel, as proposed, for example, in [1], [2], [3] for the maximum clique problem, or in [4], [5] for constraint satisfaction problems. In most cases, the decomposition is done by splitting variable domains.…”

“…In order to balance the workload on the workers, EPS ( [4], [5]) splits the initial problem in a very large number of subproblems whose resolution time can be shared by workers: all subproblems are put in a queue and workers take subproblems when they need work. Experiments in [4], [5] show us that a good decomposition is generally obtained, with EPS, by generating about 30 subproblems per worker. In [5], the average speedup reported with EPS is close to 0.5k, where k is the number of workers, on a large benchmark of instances.…”

“…A first goal of this paper is to study subproblem balance for the maximum common subgraph problem, and to show that the decomposition proposed in [4], [5] leads to very unbalanced subproblems: on our benchmark, t 0 /t max is equal to 4 when decomposing problems into 400 subproblems or so. As a consequence, the speedup cannot be higher than these bounds.…”

Parallel Strategies Selection

“…Our approach does not require to deal with a set of instances and use some sampling technique that are usually more accurate. It exploits the decomposition proposed by the embarrassingly parallel search (EPS) method recently developed [21,22]. EPS proposes to solve a problem by decomposing it into a large number of subproblems consistent with the propagation (i.e., there is no immediate failure triggered by the initial propagation of a subproblem).…”