A busy time at SC17

I have a couple of things in the conference, the first being a poster in the main conference on Tuesday about my latest developments on ePython, a very memory efficient implementation of Python for the Epiphany micro-core processor. Since I first presented ePython at SC16, we have started to look at these micro-cores far more as accelerators driven from the host. To this aim I developed an offload abstraction where the programmer can decorate functions in Python codes running on the host and for these to then be seamlessly executed on the micro-cores using ePython as the execution engine. There is also functionality to declare and manage device resident data and more information is available in this blog post.

My poster also details some benchmarking and evaluation, where I developed a neural network detecting lung cancer in 3D CT scans. Using our decorators the (computationally intensive) linear algebra kernels required for inference and training were offloaded to the micro-cores. It has worked pretty well and is competitive with other Python interpreters (such as CPython.)

On Wednesday I will give a short talk at the PRACE booth about INTERTWinE, one of the FET HPC projects that we lead in EPCC and that I work on. Specifically I will concentrate on a couple of software developments that this project has produced: a resource manager to enable efficient and fair sharing of computational resources (such as threads) between different runtimes, and our directory-cache which abstracts distributed memory spaces from task based models running over multiple nodes.

On Wednesday evening I am also involved in a BoF entitled “From outreach to education to researcher; innovative ways of expanding the HPC community” which is about public engagement and online education (such as MOOCs.)  Specifically we will be considering what the HPC community is currently doing in this area, how we can develop new approaches & technologies to support this, and how to ensure diversity.