In recent years, dynamic compilation and optimization have received a lot of
attention from many researchers as Java program and .NET runtime becomes
popular. Traditional compiler needs proper profile of program execution for
advanced compiler optimizations. Producing representative profile covering
most of input sets by itself is very difficult and sometime is not possible.
Dynamic binary re-optimization opens a way to adapt program binary according
to run-time program behaviors by collecting profile and re-optimizing binary
on the fly. Many programs have too short execution time to gather proper
profile, but they are repeatedly executed in many times. In server and
parallel applications running on multi-processors, this becomes even more
difficult because many concurrent processes cooperate and show variable
behavior on different processors in each runs.

       In this talk, I will present a framework that collects proper profile
for multi-process and multi-threaded applications and accumulates profile
information with dynamic control flow. This framework enables continuous
optimization for parallel programs and detecting abnormal behaviors even in
the single threaded program. Monitoring and profiling is non-intrusive
because it uses sampling techniques with support of hardware performance
registers. It deploys learning mechanism to adapt optimization according to
system run-time requirements such as performance, power consumption and so
on. Eventually, it will be key components of new run-time environment to
load, execute, profile and adapt program binary by itself in autonomous
fashion.