Model-Driven Comparison of State-Machine-Based and Deferred-Update Replication Schemes

Abstract

Abstract-In this paper, we analyze and experimentally compare state-machine-based and deferred-update (or transactional) replication, both relying on atomic broadcast. We define a model that describes the upper and lower bounds on the execution of concurrent requests by a service replicated using either scheme. The model is parametrized by the degree of parallelism in either scheme, the number of processor cores, and the type of requests. We analytically compared both schemes and a nonreplicated service, considering abcast-and request-executiondominant workloads. To evaluate transactional replication experimentally, we developed Paxos STM-a novel fault-tolerant distributed software transactional memory with programming constructs for transaction creation, abort, and retry. For statemachine-based replication, we used JPaxos. Both systems share the same implementation of atomic broadcast based on the Paxos algorithm. We present the results of performance evaluation of both replication schemes, and a non-replicated (thus prone to failures) service, considering various workloads. The key result of our theoretical and experimental work is that neither system is superior in all cases. We discuss these results in the paper.

Hybrid Replication: State-Machine-Based and Deferred-Update Replication Schemes Combined

“…Firstly, as our previous work [9] showed that neither the state-machine-based nor deferred-update replication scheme was superior, we were eager to combine these two into a single algorithm to bring together the best of both worlds. Secondly, we are not aware of any prior research on applying transactional semantics to state machine replication for increased expressiveness.…”

“…We added the HTR functionality to Paxos STM [9] which is an object-based DTM system that we developed to compare the SM and TR replication schemes. It builds on JPaxos [24]a highly optimized implementation of the Paxos algorithm [25].…”

“…To commit a transaction t, t's readsets, writesets and updates are broadcast to all replicas with a global order (lines [29][30]. Then, the replicas independently certify the received sets against already-committed transactions which do not precede t (lines [9][10][11][12][13][14]. If a conflict is detected, i.e., t reads data modified by a concurrent but already-committed transaction t , and neither t → t nor t → t holds, then t is aborted and restarted (line 33).…”

“…In our previous work [9], we analytically and experimentally compared the SM and TR replication models. Our results show that, surprisingly, there is no clear winner-each approach has its advantages and drawbacks, and various factors such as a workload type, parallelism on multicore CPUs, and network congestion have significant impact on performance of the SM and DU replication schemes.…”

“…To evaluate our ideas, we extended Paxos STM [9], our optimistic distributed transactional memory system, with the HTR algorithm presented in this paper. Paxos STM replicates all transactional objects (objects shared by transactions) and maintains strong consistency of object replicas.…”