
8.10 Advanced Topics 655
computing, performing the transpose in the process. Which method is better may
depend on the architecture.
Finally, synchronization can be expensive in scalable systems, so programs should
make a special effort to reduce the frequency of high-contention locks or global bar-
rier synchronization.
810 ADVANCED TOPICS
Before concluding the chapter, we cover two additional topics. The first deals with
the actual techniques used to reduce directory storage overhead in flat, memory-
based schemes. The second addresses techniques for hierarchical coherence, both
snooping and directory based.
8.10.1 Reducing Directory Storag