Por Jan Kończak (Poznan University of Technology, PL).
I will present several ways to support recovering a crashed replica in a SMR system based on MultiPaxos protocol. Typically, MultiPaxos requires to persist data on storage upon agreeing each consensus instance to provide crash-recovery. I will discuss how the persistent memory can decrease this overhead and how it can be leveraged in a SMR framework for further benefits. Then, I will show how to support recovery when the number of processes down at a time is limited. With such assumption a SMR system suffers no performance penalty from recovery support. Finally, I'll compare the proposed recovery support methods in terms of performance and recovery robustness.
Short Bio: Jan Kończak started his research work by implementing JPaxos – an efficient SMR framework based on Paxos. Then its interests shifted towards Software Transactional Memory (STM), where he investigated the use of irrevocable operations within transactions. This resulted in a couple of STM algorithms that provide both support for the irrevocable operations and never force a transaction to wait. Then he returned to studying SMR and worked on enhancing performance in crash-recovery model. In connection with his research, he spend a couple of months at EPFL in Lausanne and was an R&D intern at ARM in Cambridge.
