We have proposed two compiler-assisted software-cache schemes.
One is a page-based system (Asymmetric Distributed Shared Memory:
ADSM) which exploits TLB/MMU only in the cases of read-cache-misses.
Another is a segment-based system (User-level Distributed Shared
Memory: UDSM) which uses only user-level checking codes and
consistency management codes for software-cache.
Under these schemes, an optimizing compiler directly analyses shared
memory source programs, and performs sufficient optimization. It
exploits capabilities of the middle-grained or coarse-grained
remote-memory-accesses in order to reduce the number and the amount of
communications and to alleviate overheads of user-level checking
codes. It uses interprocedural points-to analysis and interprocedural
redundancy elimination and coalescing optimization.
We have implemented the above optimizing compiler for both schemes.
We also have implemented runtime systems for user-level cache
emulation. Both ADSM runtime system and UDSM runtime system run on the
SS20 cluster connected with the Fast Ethernet(100BASE-TX). We have
revealed that both schemes achieve high speed-up ratio with the
SPLASH-2 benchmark suite.