< BK21+ BEST Seminar Series Announcement> 

Time and Date : 16:00 ~ 17:00 Thursday 05/23/2019

Place : D604, Engineering Building #4

Title : Performance Modeling Techniques for Effective Computer System Development
Abstract:
Performance models are the main tools for computer architects to analyze the current systems and develop new systems resolving the performance issues. The speed and accuracy of performance models are of utmost importance, as they directly affect the delivery schedule and the quality of new systems. Unfortunately, the growing complexity of modern systems prohibit effective analysis and exploration of new computer system designs.

In this talk, we discuss three techniques to enhance performance models targeting various scopes of systems. First, we introduce representative stall-event analysis, which allows the architects to quickly explore the design tradeoffs of a system represented as a dependence graph (e.g., out-of-order CPU pipeline). Second, we show that focusing on unique system behaviors at right abstraction level significantly improves the speed of a full-system performance model (e.g., timing simulator). Lastly, we present a statistical modeling technique to efficiently evaluate a large-scale system (e.g., datacenter) while embracing its inherent performance variance. We also discuss future research directions to further improve the performance models and extract more valuable insights.

Presenter: Lee Jaewon, Researcher / Seoul National University

Host: Prof. Kim, Hanjun, Yonsei EEE