Institute for New Generation Computer Technology
4-28, Mita 1-chome, Minato-ku, Tokyo 108, Japan {masaki, yamauchi, chikayama}@icot.or.jp
Abstract:
PIMOS, an operating system for the parallel inference machine PIM,
provides a utility, called pool , that manages tables of
arbitrary KL1 data. This pool stores any KL1 data and enables it to be
extracted and/or referenced when needed according to a search key.
This pool is used with many application programs, reducing the cost of
program development.
However, its performance is not satisfactory for highly parallel
applications, because stored data is located in the memory of the node
where the pool was initially created. When its data is accessed from
another node, the ensuing data transfer increases network traffic, and
simultaneous access causes centralization of the computing load. To
overcome this problem, we have developed a distributed pool. The pool
is distributed to many nodes and, if a process accesses data existing
in another node, the pool caches data in its node. It provides the
same functionality as the original pool by maintaining cache
coherence. This paper describes the mechanism of this distributed
pool.
All of this proceedings (Compressed PostScript file) :
