Cutting Through the Virtual Memory Wall: From Rigid Abstractions to High-Performance, Flexible, Efficient Cooperation Between Hardware and System Software

Speaker

Konstantinos Kanellopoulos
SIGMICRO's MICRO Hall of Fame, Distinguished Artifact Award at MICRO 2023
 

Title

"Cutting Through the Virtual Memory Wall: From Rigid Abstractions to High-Performance, Flexible, Efficient Cooperation Between Hardware and System Software"

Abstract

Virtual memory (VM) is a foundational abstraction, but its rigid, legacy hardware-operating system (OS) contract has failed to keep pace with the demands of modern systems and workloads, creating a "Virtual Memory Wall" of severe performance and energy bottlenecks. This wall manifests as high address translation latency in data-intensive applications like graph analytics and high memory allocation latency in short-lived workloads like microservices. In this talk, I am going to discuss how to address these limitations by fundamentally restructuring the current hardware-OS interface to enable cooperative virtual memory mechanisms that revamp restrictive legacy abstractions. We first introduce Virtuoso, a new simulation framework that enables fast, accurate evaluation of VM components and allows identifying key VM bottlenecks. Then, I will present four new address translation schemes—Victima, TRAIL, Utopia, and Revelator—that progressively enrich the hardware-OS interface to expand translation reach, perform efficient translation prefetching, and leverage hashing to drastically improve address translation. Next, I will discuss how we tackle memory allocation overheads with Valinor, a programmable and high-performance cooperative memory allocation substrate that executes software-defined OS allocation libraries directly in hardware, completely removing costly context-switches from the critical path of execution. I will conclude by demonstrating how new simulation methodologies and architecture-level solutions can leverage hardware-OS cooperation to effectively cut through the Virtual Memory Wall, thereby enabling the high performance, energy efficiency, and flexibility critical for next-generation memory systems and workloads.

Speaker Bio

Konstantinos recently defended his PhD dissertation at ETH Zurich, advised by Prof. Onur Mutlu in the SAFARI Research Group. Before joining ETH, he received his Master’s diploma at the National Technical University of Athens (NTUA). He also spent time as a visiting student at the Huawei Zurich Research Center and Microsoft Research Cambridge. His research lies at the intersection of hardware, software, and operating systems with a focus on high performance, energy efficiency, and programmability. Lately, he has been exploring system and architecture design for Physical AI applications. His works have been published in top-tier conferences such as MICRO, ISCA, ASPLOS, HPCA, and DSN. His work has been recognized with Best Paper Awards at MICRO 2022, HiPEAC 2023, ISCA 2024, and a Distinguished Artifact Award at MICRO 2023. He also recently joined SIGMICRO's MICRO Hall of Fame.