Conflict detection via adaptive signature for software transactional memory

Transactional Memory (TM) is a promising way to coordinate concurrent threads in multi-core processors. Software transactional memory (STM) can run on conventional processors without additional hardware support. It is an option before TMs are available in commercial multi-core processors. Several TM systems use signatures to track items read and written during a transaction and to detect conflicts between transactions. Signatures summarize unbounded read/write sets in bounded bit vectors at a cost of false positives (conflicts detected when none exists). Most signatures are implemented in hardware. The size of read/write sets in different applications fluctuates in a wide range which makes it difficult to decide the size signatures. It will waste resources with long signature and cause higher false positive rate with short one. In this paper we present an adaptive signature AdapSig to detect conflicts in STM. The idea is to adapt the number and length of bit vectors to the size of read/write sets keeping an acceptable false positive rate. We implement it in STM which is different from hardware implementation. We evaluate AdapSig with a transactional benchmark suit named STAMP and improve execution time by up to 10% comparing the original one in RSTM.