High energy consumption has become a critical
problem for supercomputer systems. GPU clusters are becoming
an increasingly popular architecture for building supercomputers
because of its great improvement in performance. In this
paper, we first formulate the tasks mapping problem as a minimal
energy consumption problem with deadline constraint. Its
optimizing object is very different from the traditional mapping
problem which often aims at minimizing makespan or minimizing
response time. Then a Waterfall Energy Consumption Model,
which abstracts the energy consumption of one GPU cluster system
into several levels from high to low, is proposed to achieve an
energy efficient tasks mapping for large scale GPU clusters.
Based on our Waterfall Model, a new task mapping algorithm is
developed which tries to apply different energy saving strategies
to keep the system remaining at lower energy levels. Our mapping
algorithm adopts the Dynamic Voltage Scaling, Dynamic
Resource Scaling and β-migration for GPU sub-task to significantly
reduce the energy consumption and achieve a better load
balance for GPU clusters. A task generator based on the real task
traces is developed and the simulation results show that our
mapping algorithm based on the Waterfall Model can reduce
nearly 50% energy consumption compared with traditional approaches
which can only run at a high energy level. Not only the
task deadline can be satisfied, but also the task execution time of
our mapping algorithm can be reduced.
