Fri Aug 12, 11-12, 4750 Boelter Hall

Computing at Scale -- the Design of X10

Vijay Saraswat
IBM Watson

We present the design of the X10 programming language,
a modern object-oriented programming language
intended to express computations for high-end
computers operating with hundreds of thousands of threads.
At such large scales, data access times and latency can vary
widely across the machine, making it very difficult to support
a notion of uniform shared memory. X10 introduces the notion
of *places* to reify locality: a place contains data, together
with one or more activities operating on the data.  A computation
may have millions of places. Aggregate data-structures,
such as arrays, may be scattered across multiple places.
Activities may be spawned dynamically to operate on local
or remote data. Clocks are introduced to provide a determinate
way of managing distributed phased operations.  An activity may
wait for computation of a subactivity to terminate before progressing.
Taken together, these provide a very rich, expressive and yet
disciplined framework for concurrent programming over distributed
datastructures.  (We show, for instance, that programs using
the X10 constructs above cannot deadlock.)

Thus X10 covers those programming areas that are today handled
with a combination of OpenMP (for intra-node parallelism) and
MPI (for message-passing between nodes).  We illustrate through
several (running) examples, and discuss the status of the current
reference implementation.

Joint work with Vivek Sarkar, Kemal Ebcioglu, Christoph von Praun,
Christian Grothoff, Philippe Charles, Allan Kielstra,
Christopher Donawa and Armando Solar-lezama.

Host: Jens Palsberg