On the Relevance of Wait-free Coordination Algorithms in Shared-Memory HPC:The Global Virtual Time Case
share
High-performance computing on shared-memory/multi-core architectures could suffer from non-negligible performance bottlenecks due to coordination algorithms, which are nevertheless necessary to ensure the overall correctness and/or to support the execution of housekeeping operations, e.g. to recover computing resources (e.g., memory). Although more complex in design/development, a paradigm switch from classical coordination algorithms to wait-free ones could significantly boost the performance of HPC applications. In this paper we explore the relevance of this paradigm shift in shared-memory architectures, by focusing on the context of Parallel Discrete Event Simulation, where the Global Virtual Time (GVT) represents a fundamental coordination algorithm. It allows to compute the lower bound on the value of the logical time passed through by all the entities participating in a parallel/distributed computation. Hence it can be used to discriminate what events belong to the past history of the computation—thus being considered as committed—and allowing for memory recovery (e.g. of obsolete logs that were taken in order to support state recoverability) and non-revokable operations (e.g. I/O). We compare the reference (blocking) algorithm for shared memory, the one proposed by by Fujimoto and Hybinette <cit.>, with an innovative wait-free implementation, emphasizing on what design choices must be made to enforce this paradigm shift, and what are the performance implications of removing critical sections in coordination algorithms.
READ FULL TEXT