Minimizing Staleness and Communication Overhead in Distributed SGD for Collaborative Filtering

Distributed asynchronous stochastic gradient descent (ASGD) algorithms that approximate low-rank matrix factorizations for collaborative filtering perform one or more synchronizations per epoch where staleness is reduced with more synchronizations. However, high number of synchronizations would prohibit the scalability of the algorithm. We propose a parallel ASGD algorithm, ?-PASGD, for efficiently handling ? synchronizations per epoch in a scalable fashion. The proposed algorithm puts an upper limit of K on ?, for a K-processor system, such that per-forming ? = K synchronizations per epoch would eliminate the staleness completely. The rating data used in collaborative filtering are usually represented as sparse matrices. The sparsity allows for reduction in the staleness and communication overhead combinatorially via intelligently distributing the data to processors. We analyze the staleness and the total volume incurred during an epoch of ?-PASGD. Following this analysis, we propose a hypergraph par-titioning model to encapsulate reducing staleness and volume while minimizing the maximum number of synchronizations required for a stale-free SGD. This encapsulation is achieved with a novel cutsize metric that is realized via a new recursive-bipartitioning-based algorithm. Experiments on up to 512 processors show the impor-tance of the proposed partitioning method in improving staleness, volume, RMSE and parallel runtime.