In current large-scale distributed key-value stores, a single end-userrequest may lead to key-value access across tens or hundreds of servers. Thetail latency of these key-value accesses is crucial to the user experience andgreatly impacts the revenue. To cut the tail latency, it is crucial for clientsto choose the fastest replica server as much as possible for the service ofeach key-value access. Aware of the challenges on the time varying performanceacross servers and the herd behaviors, an adaptive replica selection scheme C3is proposed recently. In C3, feedback from individual servers is brought intoreplica ranking to reflect the time-varying performance of servers, and thedistributed rate control and backpressure mechanism is invented. Despite ofC3's good performance, we reveal the timeliness issue of C3, which has largeimpacts on both the replica ranking and the rate control, and propose the Tars(timeliness-aware adaptive replica selection) scheme. Following the sameframework as C3, Tars improves the replica ranking by taking the timeliness ofthe feedback information into consideration, as well as revises the ratecontrol of C3. Simulation results confirm that Tars outperforms C3.
展开▼
