Tensorflow for HPC

Google has developed TensorFlow, a truly complete platform for ML. TensorFlow has many ingredients, for example:

• many domain specific libraries for machine learning

• the TensorFlow domain specific data-flow language

• carefully organized input and output for data flow

• an optimizing runtime and compiler

• hardware implementations of TensorFlow operations in TensorFlow processing unit (TPU) chips

The performance of the platform is amazing, and it begs the question if it will be useful for HPC in a similar manner that GPU’s heralded a revolution.

InsideHPC created a podcast based on my lecture about this at the South African National HPC Conference CHPC 2018 Conference in Cape Town.

The slides are here and here

Performance Engineering for the SKA Telescope

PERFORMANCE ENGINEERING FOR THE SKA TELESCOPE

I gave a keynote at ICPE 2018 in Berlin about performance aspects in computing for the SKA telescope. The slides were posted here and the local copy is here.

ABSTRACT

The SKA radio telescope will be a massive world class scientific instrument, currently under design by a world wide consortium, to progress to full operation in South Africa and Australia in the mid 2020’s. The capabilities of the telescope are expected to enable major scientific breakthroughs. At the center of its data processing sits the Science Data Processor, a large HPC system with specialized software. In this lecture we will give a high level overview of the project and progress to the computing and data related architecture. Then we will discuss the work of the SDP design consortium to understand and achieve the many performance requirements leveraging hardware and algorithms. Among these is a requirement for memory bandwidth exceeding 100 PB/sec.

Dramatic Changes in HPC Storage

At the 2017 CHPC conference I gave a plenary lecture looking at developments in HPC storage past and future.

Dramatic changes for Storage and IO in HPC are upon us. Bridging several tiers of storage is becoming a necessity, software will need to improve dramatically to keep up with faster memory and storage devices, and far more nodes will be supplying storage resources. We will talk about roles for technologies such as containers, file systems, object storage and give an overview of a broad class of new data movement software and analogies with the cloud. This is an area with many exciting opportunities and presently only a handful of solutions are available or under development.

Intel has published a beautiful image outlining some of the developments

The numbers required by some of the future deployments, in particular the SKA Science Data Processor always draw attention.

The slides for CHPC IO.