Enhancing Performance Guarantees in Large-Scale Systems

Abstract

Achieving performance stability is fundamentally challenging because performance is highly impacted by application code. For example, application-defined resources such as logical queues, internal schedulers, buffer pools, and memory allocators control concurrency, buffering, and execution order. These resources are implemented entirely within application code, heavily influence execution behavior, and are invisible to the operating system.

As a result, performance degradation often comes from subtle mismanagement or contention within these internal resources. For example, when a database encapsulates heap memory into a buffer pool to cache query results, query latency depends on whether data is served from the buffer pool or fetched from disk. Contention, imbalance, or configuration choices within such resources can lead to severe and unpredictable performance degradation, even when system-level resource utilization appears normal.

In this talk, I will present my research on making application-defined resources first-class entities for performance reasoning and control. I will introduce techniques that systematically identify, monitor, and manage these internal resources to prevent catastrophic performance degradation and improve performance stability under high concurrency and dynamic workloads. These efforts aim to bridge the visibility gap between applications and the operating system, enabling principled performance reasoning and runtime mitigation in large-scale systems.

Bio

Yigong Hu is an Assistant Professor at Boston University. His research interests include computer systems, system reliability, and machine learning systems, with a focus on building reliable and efficient large-scale software infrastructure.