Costin Iancu

I am performing research in the areas of programming models and code optimization for large-scale parallel systems. I tend to favor simple and practical designs. Over the years I've been involved in multiple projects.

• Big Data Support on HPC Systems. (Intel Parallel Computing Center): There exists an incentive in the HPC community to run commercial data analytics frameworks (Hadoop, Spark) on capability systems. Adoption so far has been hampered by design decisions that are commercial data-center specific, e.g. local disk available. We plan to explore R&D for technologies to enable Spark to run on large-scale HPC systems. The IPCC is part of our larger research agenda and focuses on improving the interaction between Spark and the Lustre parallel global file system on production systems at NERSC. First notable accomplishment is that we show how to scale Spark from 100 cores to 10.000 cores on production HPC systems.
• CORVETTE (Correctness, Verification and Testing of Parallel Programs): I'm interested in tools that assist developers in bug finding or program state exploration for debugging or optimization purposes. The focus is on developing a scalable dynamic program analysis framework for the HPC programming models
• Berkeley UPC I wrote a lot of the compiler code and moved on to code optimizations, where I've been successful when using a combination of static and dynamic program analyses to guide communication and task scheduling optimizations. To guide optimizations, I'd rather use "qualitative" performance models over the more traditional approaches: track derivatives and preserve ordering rather than predict absolute values.
• THOR (Throughput Oriented Runtime): In this project, we are trying to build a software overlay network to perform on-the-fly communication optimizations for large-scale systems. There are three dimensions of control: concurrency, order, and granularity. We have developed mechanisms to regulate the concurrency of communication operations, avoid congestion, scheduling/reordering for improved throughput, etc.

The individual project pages contain more information: slides, posters, software, fun ...