Techniques to Expand Capacity in Memory Hierarchies using Compression
Licentiatavhandling, 2025

This thesis targets two open problems in the literature related to the area of data compression in memory hierarchies as a means of expanding the capacity without physically adding more memory.

For the first part, as computational demands continue to increase, hybrid memory systems that integrate High-Bandwidth Memory (HBM) as Near Memory (NM) and DRAM as Far Memory (FM) present a compelling solution for achieving high-bandwidth, large-capacity, and cost-effective main memory. However, existing flat hybrid memory architectures suffer from either excessive swap traffic or underutilized NM capacity. To address these challenges, this paper introduces HMComp, a novel flat hybrid-memory architecture that uses memory compression techniques to optimize NM usage, creating cache space for FM. By dynamically repurposing the freed-up NM capacity as a cache for FM data, HMComp effectively reduces swap traffic while maintaining full memory capacity. Additionally, a carefully designed metadata layout ensures that metadata is stored in the low-cost FM, preserving valuable NM capacity for critical application data. Experimental results demonstrate that HMComp achieves up to a 22\% improvement in single-thread performance (13\% on average) and reduces swap-related traffic by up to 60\% (41\% on average) compared to traditional flat hybrid memory systems.

The second part of this work presents COMPAT, an advanced memory compression framework designed to overcome the limitations of traditional approaches that use memory compression to expand memory capacity. Conventional memory compression frameworks often rely on a limited set of fixed compression unit sizes and struggle with misaligned block sizes relative to application data structures, resulting in suboptimal memory expansion. COMPAT introduces a novel object-based compression technique that enables fine-grained compression unit sizing without the overhead typically associated with dynamic size adjustments. The framework features a load-store unit design capable of inferring object structures with minimal changes to the memory system and no modifications to the instruction set architecture. COMPAT additionally offers fine-grain compression unit sizes and resolves the issue of compression unit size alteration overhead by a compression stabilization technique. This technique does not compress pages until the compression unit sizes of blocks in a page stabilize. Experimental results show that COMPAT achieves a 30\% improvement in compression ratio and a 35\% boost in performance over existing memory compression solutions. With the compression stabilization technique, COMPAT reduces the compression unit size alteration ratio from 22.5\% (without stabilization) to 0.9\%.

Hybrid Memory Management

Memory Compression

Object-Level Compression

EDIT-EA Lecture Hall, Rännvägen 6B, Chalmers
Opponent: Mattan Erez, University of Texas at Austin, USA



Författare

Qi Shao

Chalmers, Data- och informationsteknik, Datorteknik

COMPAT: A Compression Framework with Fine-Grain and Object-Level Compression Units

HMComp: Extending Near-Memory Capacity using Compression in Hybrid Memory

Proceedings of the International Conference on Supercomputing,;(2024)p. 74-84

Paper i proceeding

Ämneskategorier (SSIF 2025)

Datorsystem

Utgivare

Chalmers

EDIT-EA Lecture Hall, Rännvägen 6B, Chalmers

Online

Opponent: Mattan Erez, University of Texas at Austin, USA

Mer information

Senast uppdaterat

2025-03-18