A coherent shared address space provides an attractive programming
environment for parallel computing. If such an address space is to
be used in a distributed-memory system without dedicated communication
hardware, software remote-data caching mechanism (software Distributed
Shared Memory: DSM) must be used. Optimizing methods are indispensable
for improving the performance of Software DSM schemes. That is, compiler
optimization, protocol optimization, run-time system optimization, and
the interfaces that enable these optimizations are required.

We have introduced two compiler-assisted software DSM schemes as these
interfaces.
One is a page-based system (Asymmetric Distributed Shared Memory: ADSM)
that uses the TLB/MMU mechanisms only for cache-misses. The other is a
segment-based system (User-level Distributed Shared Memory: UDSM) that
uses user-level checking codes and consistency management codes for
software caching. 

We have proposed a compiler optimization framework that directly
analyzes the explicitly parallel shared-memory source programs 
and optimizes them.
The optimizing compiler exploits the capabilities of
middle-grained/coarse-grained shared-memory access to reduce the volume
of communications and to reduce the overhead for the user-level checking
codes.  It performs interprocedural points-to analysis and
interprocedural shared-access set calculations by using interval
analysis to solve redundancy elimination equations along with lazy
release consistency model.
We have implemented this optimizing compiler, Remote Communication
Optimizer :RCOP. 

We have developed the cache-coherence protocols that follow the lazy
release consistency model. Our experiment shows that SAURC (Software
emulation of Automatic Update Release Consistency) protocol provides the
best performance among the protocols that follow lazy release
consistency model. 
We have developed the lightweight run-time system for cache-coherence
based on SAURC, and we have implemented the run-time system for ADSM and
UDSM on an SS20 workstation cluster. Both systems provide high speed-up
ratios for the SPLASH-2 benchmark suite. The experimental
results show that the combination of the optimizing compiler and
Software DSM is very effective. The experimental results also show that
the performance of the ADSM scheme is limited by the communication of
unnecessary data, while that of the UDSM scheme is limited by the
instrumentation overhead.

The results of this study indicate that executing parallel shared-memory
programs with automatic optimizations for remote communications
under a general-purpose operating system on a stock network of
workstations is feasible.