enerGyPU 0.1.1-rc0

Release 0.1.1

This version of enerGyPU monitor tool is used as a submodule of DiagnoseNET application-framework for energy-workload characterization while the deep neural networks are training.

Major Features And Improvements

• Added enerGyPU_availabledevice.sh into the launcher script to get the GPU available in a multi-GPU node and pass for a high library as TensorFlow.
• Extend the energy-workload characterization in a distributed training DNN on NVIDIA Jetson TX2 capturing
  power traces with tegrastats and managing the GPU identification by each worker enerGyPU_record-cluster.sh.
• Added enerGyPU_bandwidth.sh bandwidth monitor to record between the master and each worker.

Adding enerGyPU into DiagnoseNET repository

• Clone a enerGyPUTesting branch:
  git clone -b enerGyPUTesting https://github.com/jagh/enerGyPU.git

• Build a enerGyPUTesting as submodule:
  git submodule add -b enerGyPUTesting https://github.com/jagh/enerGyPU.git

• Fix a Git detached head in a submodule:
  git checkout enerGyPUTesting

• Updated DiagnoseNet to the latest commit on enerGyPUTesting:
  git pull origin enerGyPUTesting
## Go back to DiagnoseNet directory
git add enerGyPU/
git commit -m "submodule updated"

enerGyPU 0.1.0-rc0

Release 0.1.0

This was the first version of enerGyPU monitor used with High-Performance Linpack Benchmark for recording power consumption on Multi-GPU nodes.

Major Features And Improvements

• The first level capture the GPU power traces according to whit the workload parameters and computational resources selected.
• The second level uses the enerGyPU monitor at post-processing for data visualization and statistical characterization of each computational architecture.
• The third level uses the projected multivariable machine learning models to project metrics such as, execution time, performance, power consumption and energy per watt, that could be used in accordance with workload parameters and computational resources selected.