Accelerating iterative algorithms with asynchronous accumulative updates on FPGAs

Abstract

Iterative algorithms represent a pervasive class of data mining, web search and scientific computing applications. In iterative algorithms, a final result is derived by performing repetitive computations on an input data set. Existing techniques to parallelize such algorithms typically use software frameworks such as MapReduce and Hadoop to distribute data for an iteration across multiple CPU-based workstations in a cluster and collect per-iteration results. These platforms are marked by the need to synchronize data computations at iteration boundaries, impeding system performance. In this paper, we demonstrate that FPGAs in distributed computing systems can serve a vital role in breaking this synchronization barrier with the help of asynchronous accumulative updates. These updates allow for the accumulation of intermediate results for numerous data points without the need for iteration-based barriers allowing individual nodes in a cluster to independently make progress towards the final outcome. Computation is dynamically prioritized to accelerate algorithm convergence. A general-class of iterative algorithms have been implemented on a cluster of four FPGAs. A speedup of 7× is achieved over an implementation of asynchronous accumulative updates on a general-purpose CPU. The system offers up to 154× speedup versus a standard Hadoop-based CPU-workstation. Improved performance is achieved by clusters of FPGAs.