רמי ג'וסי, הרצאה סמינריונית למגיסטר
יום שלישי, 29.7.2014, 12:30
Programmers developing GPGPU applications to run on hybrid systems are mainly facing two types of challenges,
achieving better performance and saving energy. One type of hybrid systems is recent SoC platforms. When it
comes to such integrated, SoC based systems, there is a high need to optimize for both, energy and performance.
Prior research works already suggested approaches and techniques to schedule computations for parallel execution
on the different compute devices (hence hybrid execution) with the goal of improving performance while only few
focused on energy-performance. All these research works had been targeted at hybrid platforms with CPU and discrete
GPUs, for which performance or energy can be optimized independently per device. In the SoC environment, such
optimizations are not trivial since any attempt for optimizing both energy and performance may introduce new
challenges, so that prior techniques are not applicable.
In this talk we will present a novel method to help optimizing SoC-based heterogeneous systems, to reveal the hidden
potential and achieve better energy-performance. The method aims at detecting the optimal system configuration;
such as partitioning of the computations between the compute devices in an energy aware fasion (eg. CPU and GPU), together
with enforcing resource balancing (i.e. frequencies) of each subsystem, to improve the overall energy-performance for a
given workload. Our proposed method exhibits an average of 23% improvement in energy-performance for a set of well-known
benchmarks, compared with the current HW power management methods.