One of the main obstacles to the wide adoption of transactional memory
is the interdependence between the justification for the cost of
effective architectural support for TM and the performance impact of
TM on important applications. Without showing clear positive impact of
TM on the performance of important applications, it is difficult to
justify architectural support of TM that goes beyond small
transactions. On the other hand, software-only TM implementations have
not delivered robust performance that can motivate application
developers to adopt the TM programming model. After years of wide
interest in TM, applications that can be clearly identified as TM
killers apps remain elusive.  In this talk, we explore the
characteristics that would make an application a candidate for being a
TM killer app, and what pitfalls reduce the value of TM to some
applications. Three main application characteristics are identified as
critical to identification as a TM killer app: (1) importance of
performance, (2) inherent high concurrency, and (3) irregular
multi-object transactional span.

A key purpose of TM is enabling concurrency. Therefore, the
performance of a killer app must be important enough to warrant
aggressive parallelization. TM enables concurrency but does not create
it. Therefore, a killer app must have inherent
concurrency--irrespective of concurrency limitations related to the
use of specific synchronization such as lock contention.  The
importance of application performance combined with the high overheads
of TM may justify investment in low-overhead fine-grain
synchronization by expert programmers. Therefore, a TM killer app must
be very difficult to parallelize effectively and efficiently without
TM.  Transactions that span multiple data objects with irregular
access patterns present a nearly impossible challenge to expert
programmers to develop implementations that are both efficient and
maintainable. The importance of code maintainability offered by TM due
to its composability trump the performance advantage of fine-grain
synchronization.