Clustered Checkpointing and Partial Rollbacks for Reducing Conflict Costs in STMs

Abstract

A Software Transactional Memory is a concurrency control mechanism that executes multiple concurrent, optimistic, lockfree, atomic transactions, thus alleviating many problems associated with conventional mutual exclusion primitives such as monitors and locks. With the advent of massive multi-cores, more transactions can be initiated concurrently, however resulting in an increase in the percentage of conflicting transactions. Each time a transaction conflicts, it imposes a significant cost on the system, originating from the need to abort and redo all the operations, including the costly shared memory read operations, thus making the overall system significantly heavy and impractical. We present an algorithm, Clustered Checkpointing and Partial Rollback (CCPR), for reducing the conflict costs of transactions in the face of increasing conflicts. The algorithm is based on intelligent checkpointing of transactions as they proceed, and, in case of conflict, rolling them back to a safe, consistent, intermediate checkpoint, thus reducing conflict costs. The intelligence of the algorithm lies in the fact that as conflicts decrease, the checkpointing costs go low, however, when conflicts increase, the checkpointing costs increase but are still pretty much less than the amount of savings obtained by the partial rollback of the conflicting transactions. We simulated several applications in the CCPR framework and found that it can result in as good as 17% reduction in the conflict costs originating from the need to redo all the shared memory read operations.