Design And Implementation Of Reader-Preference Lock Mechanism With RDMA

In recent years,distributed systems based on Remote Direct Memory Access(RDMA)have become increasingly more common.As a key component of the system,distributed lock is used to implement concurrency control and operation serialization while accessing shared resources.Lock usually guarantees first-come-first-served scheduling.But in this way the long-latency write leads subsequent reads to significantly-long wait and limits the concurrency of reads.Note that in the flexible consistency scenarios,such as the distributed machine learning framework “parameter server”,it is not necessary to update and read the parameters in order strictly,because multiple iterations ensure the correctness of the program even if part of reads and writes are reordered.Aiming at the above scenarios,in order to improve the overall performance of the system,a RDMA-based schedulable reader-preference lock mechanism is proposed,which allows a write to be reordered with following reads and avoids the blocking caused by the write.The lock mechanism has the following characteristics.Firstly,the RDMA fetch_and_add(FA)verb is used to implementing the lock representation,acquisition and release.And lock operations don't invlove CPU in server side.Secondly,The use of the RDMA primitive without notification reduces the polling overhead and the communication delay.Thirdly,it achieves a fair scheduling strategy without global queue knowledge,guaranteeing the maximum waiting time of each node by limiting the number of reads scheduling before write.Finally,the timeout mechanism is introduced to solves the problem that the lock cannot be released when the node fails or the network fails.And the node who acquires the lock maintains the lock expire time and performs timeout processing.Based on this mechanism,exclusive write locks and shared read locks are implemented,And a shared memory system based on the mechanism is implemented.At the same time,three comparison schemes are implemented: baseline,DrTM and DSLR.Baseline uses RDMA message semantics and guarantee first-come-first-served scheduling;DrTM and DSLR both uses RDMA atomic primitives,DSLR guarantees first-come-first-served scheduling while DrTM doesn't guarantee it.The above schemes are tested for quantitative performance analysis using the load generated by YCSB(Yahoo! Cloud Serving Benchmark).The experimental results show that the throughput of the mechanism(the number of locks acquired per second)is 90% higher than the baseline.Under the read-write balanced load,the throughput increased by 10% compared to the latest study DSLR,which indicates that the reader-preference strategy is effective.