A space-efficient on-chip compressed cache organization for high performance computing

• Published in: Lecture notes in computer science pp. 952 - 964
• Main Authors: Yim, Keun Soo, Lee, Jang-Soo, Kim, Jihong, Kim, Shin-Dug, Koh, Kern
• Format: Conference Proceeding Book Chapter
• Language: English
• Published: Berlin, Heidelberg Springer-Verlag 01.01.2004; Springer Berlin Heidelberg; Springer
• Series: ACM Conferences
• Subjects: Applied sciences; Computer science; control theory; systems; Computer systems and distributed systems. User interface; Exact sciences and technology; fine-grained management; internal fragmentation problem; on-chip compressed cache; Parallel processing; processor-memory performance gap; Software; fine-grained management; internal fragmentation problem; parallel processing; processor-memory performance gap; on-chip compressed cache; High performance; Parallel algorithm; Table; Cache memory; Metadata; Distributed computing; Memory organization; Simulator; Fragmentation
• ISBN: 9783540241287; 3540241280
• ISSN: 0302-9743; 1611-3349
• DOI: 10.1007/978-3-540-30566-8_109

In order to alleviate the ever-increasing processor-memory performance gap of high-end parallel computers, on-chip compressed caches have been developed that can reduce the cache miss count and off-chip memory traffic by storing and transferring cache lines in a compressed form. However, we observed that their performance gain is often limited due to their use of the coarse-grained compressed cache line management which incurs internally fragmented space. In this paper, we present the fine-grained compressed cache line management which addresses the fragmentation problem, while avoiding an increase in the metadata size such as tag field and VM page table. Based on the SimpleScalar simulator with the SPEC benchmark suite, we show that over an existing compressed cache system the proposed cache organization can reduce the memory traffic by 15%, as it delivers compressed cache lines in a fine-grained way, and the cache miss count by 23%, as it stores up to three compressed cache lines in a physical cache line.