HTMT-C: Programming A Petaflop Scale Machine

Ryan, Amaral, Ruiz, Gao



ABSTRACT

Point design studies for future high performance computers anticipate
machines that incorporate new and non-conventional technologies. Such
machines will be constructed with a large number of processors, use deep
memory hierarchies and exhibit variable latencies. Architectures and
programming models must cope with these characteristics. In this paper we
present an explicitly multithreaded language developed for the Hybrid
Technology Multi-Threaded (HTMT) project. This language implements a new
programming model called {\em percolation}. The percolation model, 
designed to manage data movement within machines that have deep 
memory hierarchies and unpredictable latencies, requires that
instructions of a program segment be paired with the data required
by these instructions before they are shipped to the vicinity of the
fastest processors in the machine. Under this model, the programmer
or compiler may explicitly control the movement of data within
the machine's memory hierarchy.

We will also briefly describe the HTMT architecture with its
supreconducting processors, processor in memory technology, fast
optical networks, and holographic memory technology.