# Summarise Contribution & Motivation

- Energy efficiency is a primary design goal for all systems
- Power management not exposed to end user
- This makes it difficult to design general purpose runtime to schedule work between the CPU and GPU
- Black box approach is used where power model is computed once for each processor
- This is done because users don't have access to DVFS (dynamic voltage and frequency scaling)
- Distributes work across CPU and GPU in such a way to optimise for power metric

# Methodology

- Automatically characterising behaviour is non trivial.
- Measure cpu cores, gpu cores, ring interconnect and LLC = package power
- Probed processor power use under various workloads to model power consuptions
- Used microbenchmarks to get power characterisations
- Application is profiled too to determine memory or compute bound
- Use a mathematical algorithm to produce gpu offload ratio to minimise a given metric
- To implement, they use Concord framework, heterogeneous C++ but algorithm is not tied
- Work stealing from CPU
- Ran a number of different benchmarks to evaluate performance

# Critical Assessment

- Does well by using black-box approach
- Other solutions relied on static analysis or existing knowledge of processor
- Their solution is run at user level and doesn't require controlling of DVFS
- Limited to applications that benefit from offloading work to GPU