Research

SCoOL – Scalable Common Optimization Library

SCoOL is a simple programming model designed to facilitate and accelerate the search space explo- ration phase of the optimization processes. The main premise of the model is to allow the end-users to focus on the optimization details (e.g., objective function, constraints for the search space, direction in which search space should be explored, etc.), and to delegate the exploration of the resulting search space to an efficient runtime system. The runtime system itself is abstracted from the end-user, and is akin to the classic Bulk Synchronous Parallel (BSP) model. As in BSP, the exploration proceeds in super-steps, where each super-step is distributed across many processing elements running in parallel. However, SCoOL introduces two additional concepts. First, it allows the end-users to specify and maintain a globally shared state that can be used to track, in a somewhat consistent way, an auxiliary information (for example, about the best discovered solution thus far). Second, it introduces a concept of partitioner. Partitioner is specified by the end-user, and provides hints about how to distribute the search space between processing elements to maximize locality (e.g., the neighboring portions of the search space may be handled by the same element to reduce potential communication overheads due to data exchange between processing elements). SCoOL provides several ready-to-use runtime implementations that target shared memory systems as well as distributed memory clusters.

SABNA – Scalable Accelerated Bayesian Network Analytics

SABNA is an open source software suite of efficient algorithms for exact (i.e., globally optimal) Bayesian networks learning. The main idea behind the toolkit is to combine various data optimization techniques and advanced parallel algorithms to achieve scalable implementations capable of processing data instances with hundreds of variables.