A comparison of task pools for dynamic load balancing of irregular algorithms

Abstract

SUMMARYSince a static work distribution does not allow for satisfactory speed-ups of parallel irregular algorithms, there is a need for a dynamic distribution of work and data that can be adapted to the runtime behavior of the algorithm. Task pools are data structures which can distribute tasks dynamically to different processors where each task specifies computations to be performed and provides the data for these computations. This paper discusses the characteristics of task-based algorithms and describes the implementation of selected types of task pools for shared-memory multiprocessors. Several task pools have been implemented in C with POSIX threads and in Java. The task pools differ in the data structures to store the tasks, the mechanism to achieve load balance, and the memory manager used to store the tasks. Runtime experiments have been performed on three different shared-memory systems using a synthetic algorithm, the hierarchical radiosity method, and a volume rendering algorithm.

Analytical Modeling of Pipeline Parallelism

“…For instance, load balancing schemes using tasks queues on shared-memory architectures are described in [11] and [12]. Languages such as Cilk [3], and X10 [21] have proposed efficient implementations of dynamic scheduling based on work stealing [4].…”

Executing task graphs using work-stealing

“…For example, Hoffman, Korch, and Rauber [10], [13] describe and empirically evaluate a variety of implementations of task pools in software and hardware. They focus on the case where tasks have hierarchical dependencies, i.e., a parent task depends only on child tasks that it creates.…”

Scalability of Macroblock-level Parallelism for H.264 Decoding

“…The thread pool consists of a group of worker threads that wait for work on the task queue [13]. The dependencies of each MB are expressed in a dependence table.…”

“…Thread synchronization for task pools can be implemented in software or hardware. Software implementations uses blocking locks (like P-threads mutex), spin-locks or non-blocking synchronization [13]. Thread synchronization by hardware consists on implementing of the basic operations of the task pool, like enqueue and dequeue, using hardware controlled data structures [18], [19].…”