Snake: Control Flow Distributed Software Transactional Memory

Remote Method Invocation (RMI), Java’s remote procedure call implementation, provides a mechanism for designing distributed Java technology-based applications. It allows methods to be invoked from other Java virtual machines, possibly at different hosts. RMI uses lockbased concurrency control, which suffers from distributed deadlocks, livelocks, and scalability and composability challenges. We present SnakeDSTM, a distributed software transactional memory (D-STM) that is based on the RMI as a mechanism for handling remote calls and transactional memory for distributed concurrency control, as an alternative to RMI/locks. Critical sections are defined as atomic transactions, in which reads and writes to shared, local and remote objects appear to take effect instantaneously. The novelty of Snake-DSTM is in manipulating transactional memory by moving control to remote nodes, rather than remote nodes’ data being copied to the node at which the transaction runs. Transaction metadata is detached from the transactional context, and the dynamic two phase commitment protocol (D2PC) is employed to coordinate the voting process among participating nodes toward making distributed transactional commit decisions. We propose a simple programming model using (Java 5) annotations to define critical sections and remote methods. Instrumentation is used to generate code at class-load time, which significantly simplifies user-space end code. No changes are needed to the underlying virtual machine or compiler. We describe Snake-DSTM’s architecture and implementation, and report on experimental studies comparing it against competing models including RMI with mutual exclusion and read/write locks, distributed shared memory (DSM), and dataflow-based D-STM. Our studies show that Snake-DSTM outperforms competitors by up to 12× on different workloads using a 120-node system.