Recent work in replicated datastores has focused on making availability and low latency the primary requirements. We present two directory-based data- stores which make strong application semantics a primary requirement based on ideas from cache coherence in multiprocessors and distributed shared memory. The two datastores are implemented and evaluated against the Apache Cassandra distributed datastore in a simulated geo-distributed envi- ronment with workloads generated by the Yahoo! Cloud Serving Benchmark. We find that while there are cases where our systems perform with better average latency, the variability in the request latencies is undesirable due to expensive synchronization operations.
【 预 览 】
附件列表
Files
Size
Format
View
A lazy directory-based implementation of consistent replicated storage