CppCon 2021

Writing programs that run on clusters of computers is notoriously tedious and error-prone, particularly for applications like genomics, sparse linear algebra, and graph analytics, which often require fine-grained, irregular communication.

In this talk, I'll argue that the Partitioned Global Address Space (PGAS) programming model, which allows remote access to a portion of the memory space on remote processes in a distributed program, is an effective model for implementing portable, high-level C++ data structures that automatically distribute data, while taking advantage of the high-performance RDMA network primitives available in most supercomputer and data center environments.

I'll discuss how our distributed data structures library BCL builds upon the concept of a global pointer to build a collection of high-performance, high-level data structures, including hash tables, queues, and dense and sparse matrices, that can be used across a variety of communication libraries, including MPI, OpenSHMEM, and GASNet-EX. I'll touch on the some of the important issues that arise in implementing distributed data structures, such as synchronization, remote reference objects, and support for GPUs.