Speculations: providing fault-tolerance and recoverability in distributed environments
Cristian Ţăpuş, Jason J. Hickey · Hot Topics in System Dependability · 2006
Building safe and reliable programs is an important but difficult endeavor. The challenge is even greater in the context of distributed environments, which may involve complex synchronization operations in the presence of process and network failures. Transactions are one of the earliest and simplest abstractions for reliable concurrent programming [2]. They provide fault-isolation by guaranteeing the atomicity, the consistency and the durability of the actions performed as part of the transaction. Traditional transactions also provide isolation, which prevents the independent actions inside of a transaction from being visible to the outside world until the transaction either aborts or commits. In this paper we consider the case where multiple processes may cooperate in a transaction, using message passing for communication. We relax the transactional isolation property to permit inter-process communication while executing inside the transaction. This model can improve performance and provide fault-tolerance for distributed applications. We call these transactions with relaxed isolation speculations, and we introduce them as programming language primitives. Traditional checkpointing and rollback mechanisms used to provide recoverability are also similar to our approach. However, there are a few differences, as follows. Speculations can provide programs with alternate execution paths upon rollback. Speculations are lightweight checkpoints that are stored in memory and can be coupled with real checkpointing mechanism for increased reliability. Speculations are exposed as programming language primitives that have a semantics closer to that of transactions than that of checkpoints. Our system adapts mechanisms designed for checkpointing/rollback systems [1] to ensure safe recovery lines in case of distributed speculation rollback. While speculations are similar to the concept of lookahead-rollback introduced by the TimeWarp [3] mechanism, we extend the concept by allowing both explicit and implicit speculations through programming language extensions. The main contributions of this paper include: (1) the introduction of a new programming model based on speculations, (2) the definition of new speculative programming language constructs for distributed applications, (3) the description of a prototype implementation of speculations in the Linux kernel where speculative operations, including distributed commit and rollback, are transparent.