Aspire

Abstract

Many vertex-centric graph algorithms can be expressed using asynchronous parallelism by relaxing certain read-after-write data dependences and allowing threads to compute vertex values using stale (i.e., not the most recent) values of their neighboring vertices. We observe that on distributed shared memory systems, by converting synchronous algorithms into their asynchronous counterparts, algorithms can be made tolerant to high inter-node communication latency. However, high inter-node communication latency can lead to excessive use of stale values causing an increase in the number of iterations required by the algorithms to converge. Although by using bounded staleness we can restrict the slowdown in the rate of convergence, this also restricts the ability to tolerate communication latency. In this paper we design a relaxed memory consistency model and consistency protocol that simultaneously tolerate communication latency and minimize the use of stale values. This is achieved via a coordinated use of best effort refresh policy and bounded staleness. We demonstrate that for a range of asynchronous graph algorithms and PDE solvers, on an average, our approach outperforms algorithms based upon: prior relaxed memory models that allow stale values by at least 2.27x; and Bulk Synchronous Parallel (BSP) model by 4.2x. We also show that our approach frequently outperforms GraphLab, a popular distributed graph processing framework.

STRONGHOLD: Fast and Affordable Billion-Scale Deep Learning Model Training

“…Dorylus and other memory [58] or computation [59] optimization techniques can be used in combination with STRONGHOLD to utilize low-cost CPU threads to train GNNs. Furthermore, STRONGHOLD can also be used together with asynchronous training [60] to further reduce the waiting time across training epochs, but care must be taken to avoid slowing down model convergence [61].…”