Para-ksm - Parallelized Memory Deduplication with Data Streaming Accelerator

Explore a 17-minute conference presentation that introduces Para-ksm, a novel approach to parallelized memory deduplication using Data Streaming Accelerator (DSA) technology. Learn how researchers from the University of Illinois Urbana-Champaign, Daegu Gyeongbuk Institute of Science and Technology, and Yonsei University address the significant performance overhead of Kernel Same-page Merging (ksm) in datacenter environments. Discover how traditional ksm creates a substantial "datacenter tax" that degrades co-running application performance by 1.6–5.8×, limiting its deployment despite the rising costs of server memory. Understand the initial approach of DSA-ksm, which offloads page comparison and checksum computation from CPU to the Data Streaming Accelerator, reducing performance degradation to 1.0–1.6× but suffering from lower deduplication rates due to long offloading latencies through on-chip PCIe. Examine the innovative Para-ksm solution that redesigns the memory deduplication process to exploit DSA's batching capabilities, enabling operations on multiple pages per offload rather than single pages. Analyze the performance improvements achieved by Para-ksm, which increases memory deduplication efficiency per CPU cycle by 31–50% while maintaining performance degradation at only 1.3–2.7×, making it a more viable solution for widespread datacenter deployment.