Text

theses

Virtual memory offers a simple hardware abstraction to programmers freeing them from the tedious process of manual memory management. However, the emergence of new memory technologies is posing challenges for conventional virtual memory. Homogeneous memory systems are being replaced by complex heterogeneous systems with multiple memory devices with different latency, bandwidth, and capacity characteristics. This poses two problems. The first is that operating systems (OSes) must migrate pages among the heterogeneous memory devices based on attributes like page hotness and proximity to the compute unit/accelerator that uses the data. As this thesis shows, current support for page migration is infeasibly slow on emerging hardware, both due to the slow speeds of data movement and metadata update operations like TLB shootdowns. The second is that the ever-increasing aggregate capacities of these emerging heterogeneous memory systems pose immense pressure on TLBs, aggravating address translation overheads. This thesis addresses these problems by proposing modest hardware/software techniques that achieve a more efficient virtual memory system via fast hardware support for translation coherence, software support for faster page copies, and hardware/software co-design that compresses TLB entries to reduce address translation overheads.

Page migration is the means by which OSes can dynamically shift data to the memory devices that best benefit latency, bandwidth, capacity, and persistence characteristics in different phases of the program lifetime. The key to good performance is fast page migration. This thesis attacks two bottlenecks that currently constrain page migration performance --- high-overhead translation coherence and low-throughput page copying. To mitigate the first source of overhead, this thesis implements hardware support for translation coherence by fusing it with existing cache coherence protocols. To mitigate the second source of overhead, this thesis implements OS support that parallelizes, aggregates, and consolidates page migration operations to maximize migration throughput.

Heterogeneous systems are also continuing a trend that has long been seen with traditional homogeneous memory systems --- the drive towards ever-increasing memory capacities. Specific types of emerging systems with die-stacking technologies and byte-addressable persistent memories are further accelerating the total physical memory capacity that must be addressable for each process. Consequently, page tables are becoming bigger and TLB misses more frequent. To mitigate increasing address translation overheads, this thesis offers software techniques to facilitate the possibility of compressing TLB entries which rely on translation contiguity.

In summary, this work upgrades virtual memory to effectively support heterogeneous memory systems with high-performance page migration and scalable address translation. In so doing, this dissertation identifies bottlenecks in the existing virtual memory, profiles the performance impacts of these bottlenecks, and proposes hardware and software solutions to remedy them.

ETD_9427

Ph.D.

Includes bibliographical references

by Zi Yan

doi:10.7282/t3-gyad-ga14

ETD doctoral

The author owns the copyright to this work.