Ph.D Thesis

Ph.D StudentRotem Efraim
SubjectHigh Performance Computing in Physically Constrained
DepartmentDepartment of Electrical and Computers Engineering
Supervisors PROFESSOR EMERITUS Ran Ginosar
PROF. Avi Mendelson
Full Thesis textFull thesis text - English Version


Moore's law has been a powerful driving force in the computer industry for over four decades allowing increased transistor density and compute performance. Dennard scaling allowed this increase in transistor density while maintaining the power roughly fixed.  At sub-micron feature size, Dennard scaling reached its limits. Improvements in process technology continue to provide higher transistor density but power, thermal and energy constraints have become major obstacles. The increased transistor density is used to integrate more cores on the same die, either in a symmetric chip multi-processing (CMP) or asymmetric architectures. Such architectures deliver more compute performance within the power constraints.

Our research in this context attempts to answer the question, how can we best architect and manage a computer system in order to maximize user experience within the various power constraints?  

We studied the following power constraints:

Hierarchy of power delivery:

Power delivery is a major challenge for high end computers. We introduced a novel clustered multi-core power delivery system, along with management and scheduling techniques that maximize compute performance under sustained power delivery constraints. We further addressed instantaneous power excursions and developed novel compiler-assisted power management techniques to overcome them.  

Power and thermals:  

Much research has studied CPU junction temperature management. Recent trends of small form factors, such as smart phones, tablet computers and UltraBook?, introduce new power and thermal management challenges. We introduce system skin temperature as an additional thermal limiter and study the implications of power consumption profiles on compute performance.


Total system energy is another well-understood system constraint. Two approaches dominate the research and industry. One assumes that the CPU dominates the total system energy consumption; lower voltage and frequency therefore yield lower energy. Small cores or heterogeneous architecture further reduce energy under this assumption. The opposite approach claims that other system components dominate and therefore it is better to run fast and go to idle state as quickly as possible. Our novel H-EARtH algorithm demonstrates that an intermediate frequency point is often more efficient and implements an optimal frequency calculation at run time. We also demonstrate that small cores in heterogeneous systems do not always offer energy savings and the proposed H-EARtH algorithm can select the right core for the right workload. This approach was extended beyond the individual system, to a total data center energy management.