M.Sc Thesis

M.Sc StudentKosyakovsky Oleg
SubjectApproaches to Managing Trace Cache in Computer Systems
DepartmentDepartment of Electrical and Computers Engineering
Supervisors PROF. Avi Mendelson


This work proposes a new technique to improve the performance and power consumption of trace-cache architectures. The new technique uses program profiling and is based on the observation that traces can be classified into four different classes in respect to their appearance patterns. We name these classes after different kinds of travelers: rare traveler is a trace which appears infrequently and is used very little each time; seasonal tourist appears infrequently but shows a lot of activity during a few appearances; frequent flyer is a trace that gets activated repeatedly throughout the program and tends to be re-executed extensively each time it is brought to the trace cache; air crew is a trace being used frequently during most of the program lifetime.

In the current technology, it is not worth to build a trace that does not show locality of references, since it wastes more performance and power than the gain from using trace-cache. Thus, we propose to use the above classification as a guideline for the processor’s front-end regarding when to build a trace, and when to fetch it directly from the instruction cache.  This work will show that the new mechanism can be very effective for managing trace caches that suffer from high miss rate, in particular small to medium trace caches. Even more, it’s shown that our mechanism could be an enabler for some of those trace-cache sizes that aren’t beneficial at all with the standard LRU scheme of trace-cache management.