Parallel Programming with Global Asynchronous Memory: Models, C++ APIs and Implementations

In the realm of High Performance Computing (HPC), message passing has been the programming paradigm of choice for over twenty years. The durable MPI (Message Passing Interface) standard, with send/receive communication, broadcast, gather/scatter, and reduction collectives is still used to construct parallel programs where each communication is orchestrated by the developer-based precise knowledge of data distribution and overheads; collective communications simplify the orchestration but might induce excessive synchronization. Early attempts to bring shared-memory programming model—with its programming advantages—to distributed computing, referred as the Distributed Shared Memory (DSM) model, faded away; one of the main issue was to combine performance and programmability with the memory consistency model. The recently proposed Partitioned Global Address Space (PGAS) model is a modern revamp of DSM that exposes data placement to enable optimizations based on locality, but it still addresses (simple) data-parallelism only and it relies on expensive sharing protocols. We advocate an alternative programming model for distributed computing based on a Global Asynchronous Memory (GAM), aiming to avoid coherency and consistency problems rather than solving them. We materialize GAM by designing and implementing a distributed smart pointers library, inspired by C++ smart pointers. In this model, public and private pointers (resembling C++ shared and unique pointers, respectively) are moved around instead of messages (i.e., data), thus alleviating the user from the burden of minimizing transfers. On top of smart pointers, we propose a high-level C++ template library for writing applications in terms of dataflow-like networks, namely GAM nets, consisting of stateful processors exchanging pointers in fully asynchronous fashion. We demonstrate the validity of the proposed approach, from the expressiveness perspective, by showing how GAM nets can be exploited to implement higher-level parallel programming models, such as data and task parallelism. As for the performance perspective, the execution of two non-toy benchmarks on a number of different small-scale HPC clusters exhibits both close-to-ideal scalability and negligible overhead with respect to state-of-the-art benchmark implementations. For instance, the GAM implementation of a high-quality video restoration filter sustains a 100 fps throughput over 70%-noisy high-quality video streams on a 4-node cluster of Graphics Processing Units (GPUs), with minimal programming effort.