Energy Efficient Computing Using Static-Dynamic Co-optimizations

摘要: Energy consumption is a primary concern of current day computing systems -- from handheld battery operated to servers in data centers connected wall power. Research academia as well industry has focused on variety techniques for minimizing energy while maintaining good level performance. The most effective combine the best static (or offline) and dynamic online/runtime) optimizations obtain solution. Static can be more complex afford take global view application/computation being executed system however, such have conservative nature because they cannot anticipate all different scenarios that appear at runtime. Dynamic information aboutthe given input since low overhead, only local computation/ application complexity optimization relatively low. An additional direction needs considered determine whether optimizationsshould implemented software or hardware. In this thesis, I present study three aspects efficient use combination minimize consumption. first aspect consider variability execution time applications during scheduling voltage frequency scaling (DVFS) capable yet maintain desired main idea construct schedule table offline perform simple look-up second application-level reliability tolerate certain error outputs. profile-guided compilation strategy identify critical instructions monitoring technique handle corner cases Finally, third investigate flexibility customizing instruction sets processors improve efficiency. benefits compiler directedoptimizations generating custom which are within modified processor pipeline an architecture mechanism detecting roll-back safe state. Additionally, set dynamically