Supporting Transactional Memory in Distributed Systems based on the Data-Flow Model

A distributed TM model supports the TM API in a distributed system consisting of a network of nodes that communicate by message-passing links. Supporting TM in distributed systems is motivated by the difficulties of lockbased synchronization methods employed by existing distributed control-flow programming models such as RPCs. We consider Herlihy and Sun’s data-flow distributed TM model [3]. In this model, transactions are immobile (running at nodes which invoke them), but objects move from node to node. Transactional synchronization is optimistic: a transaction commits only if no other transaction has executed a conflicting access. The core of the design of a distributed TM system is composed of two elements. The first element is the conflict resolution strategy. Two transactions conflict if they access the same object and one access is a write. Most existing TM implementations adopt a conflict resolution strategy that aborts one transaction whenever a conflict occurs—e.g., a contention management module [2]. The second element is the distributed cache-coherence protocol. When a transaction attempts to access an object in the network, the distributed cache-coherence protocol must locate the latest cached copy of the object, and move a readonly or writable copy to the requesting transaction. The cache-coherence protocol must guarantee that at any time, there exists only one writable copy of each object in the system.