NOTE: THIS TALK IS TAKING PLACE IN ROOM 2311 IN OLD CS.

MADNESS (Multiresolution Adaptive Numerical Environment for
Scientific Simulation) which received an R&D 100 Award in 2011,
is a general purpose numerical framework for fast computation
with guaranteed precision in many dimensions. It is portable
from laptops to the largest supercomputers, including the Cray XE
and IBM Blue Gene/Q systems, and is open-source under GPL2 with
active developers/users at multiple universities and DOE
laboratories across the USA, Europe, Japan, and the UK. Prior
support from NSF, DOE, and has resulted in the MADNESS parallel
runtime being a powerful environment for the composition of a
wide range of scalable algorithms beyond its original
application.

The central elements of the parallel runtime are a) futures for
hiding latency and managing dependencies, b) global namespaces
with one-sided access so that applications are composed using
names or concepts central to the application, c) non-process
centric computing through remote method invocation in objects in
global namespaces, d) dynamic load balancing and data
redistribution, and e) a task-based execution model optimized for
efficient execution on multi-core processors. Using a
continuation-passing style to extend map-reduce, execution has
been demonstrated on GPGPUs. Internally, MADNESS employs a
dataflow model by expressing dependencies using futures to pass
results of still incomplete or pending tasks (local or remote) as
arguments of new tasks that are scheduled for execution once all
dependencies are satisfied.

In this seminar, we will review the current MADNESS parallel
programming environment and runtime, examining both successes
and challenges. We will also explore the emerging TESSE
that is emerging from an NSF-funded collaboration with the
PaRSEC team at the University of Tennessee, Knoxville
and the TiledArray team at Virginia Tech.