Live Webcast 15th Annual Charm++ Workshop

MPI's support for parallel libraries is based on partitioning the underlying system across ranks and time. As applications grow to encompass more components and work more dynamically, partitioning will become progressively less productive and scalable. The Charm++ parallel programming system offers an alternative in its support for asynchronous adaptive overlap of independent components. In this talk, I describe a few libraries I have worked on in Charm++, ranging across parallel I/O, dense linear algebra, and shared distributed-memory arrays. I will discuss application usage, performance results, and some future directions that this body of work points to.