Transactional Memory for Reliability

It is foreseen that technology trends will increase the transient and permanent fault rates in future processors. Thus providing reliability for both the applications running on personal computers and running on mission-critical systems is becoming an absolute necessity. A reliable system requires the inclusion of two key capabilities: 1) error detection and 2) error recovery mechanisms. Transactional Memory (TM) provides an ideal base for both error detection and error recovery. First, TM provides mechanisms to abort transactions in case of a conflict, thus they discard or undo all the tentative memory updates and restart the execution from the beginning of the transaction. Thus, a transaction’s start can be viewed as a locally checkpointed stable state which can be used for error recovery. Second, transactional semantics allows the error detection to be deferred until a transaction commits (or the value becomes externally visible), so that the cost of error detection can be reduced compared to traditional error detection schemes (in which error detection is conducted et every instruction [26]) while its efficiency can be increased.