Dr. Konstantinos Parasyris

Lawrence Livermore National Laboratory

Abstract

As we approach the limits of Moore’s law, researchers are exploring new paradigms for future high-performance computing (HPC) systems. Approximate Computing (AC) has gained traction by promising to deliver substantial computing power. However, approximate computing faces numerous challenges.

Some of the challenges of AC include: 1) How does the developer define which code should be approximated. Selecting and applying an Approximation Technique (AT) to codes is a challenge. There exist multiple ATs each with different constraints, errors, and performance characteristics. The AT selection defines performance gains obtained with respect to the original construct execution time, and the induced approximated construct output error. 2) Where to approximate. Developers cannot blindly apply an AT on any construct in the application. Since developers do not always know how approximation errors in these constructs affect the application's output quality. So, developers require methods to guide them and apply ATs on code constructs that when approximated are likely to introduce acceptable errors. In this talk, I will focus on presenting tools that allow the user to address both challenges.

Biography

Dr. Parasyris is a computer scientist at Lawrence Livermore National Laboratory in the CASC division. His Ph.D. focused on system software techniques to enhance the reliability of modern platforms. He continued researching fault-tolerance techniques during his post-doctoral studies at Barcelona Supercomputing Center. Briefly, the goal of his research is to identify opportunities to disrupt the power/QoS/performance balance of current computing HPC systems, towards lower energy computation without performance loss and with the quality of results within a user-specified tolerance