<b>As Moore's law is coming to a slow but sure end, scientific computing faces many challenges. Some additional challenges are also arising from the fact that the scientific community generally, and quantum chemistry and electronic structures community in particular, is solving increasingly ambitious, complex and large computational problems using increasingly sophisticated methods. These sophisticated methods are typically harder to parallelize, at least when high parallel efficiency is required. We think that adaptive run-time systems hold the key to high performance and programmer productivity in this era. Such systems can observe the application behavior while the program is executing, and monitor the machine behavior as well; further, they can take actions to optimize application execution during runtime. Of course, to accomplish this, they must have access to some levers in the application. In Charm++, which is a C++ based parallel programming system, a computation is expressed as an interacting collection of objects, and the adaptive runtime system mediates scheduling and communication. In addition it can also migrate objects across nodes of the system. With the degrees of freedom so generated, Charm++ runtime system can do complex optimizations such as dynamic load balancing, power energy and temperature optimizations, fault tolerance, wide supporting automatic overlap of communication and computation to maximize the utilization of the network. I will describe the basic ideas, and illustrate their utility on the variety of applications including NAMD and OpenAtom.</b>

Charm++: Necessity of adaptive runtime systems for programming future parallel computers illustrated with NAMD and OpenAtom

255th National Meeting (2018)

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.