Sunday, September 27, 2026  ·  8:00–10:00 AM     Walnut     2 hours, in person

Presenters: Binuraj Ravindran, Principal Engineer, Intel Corporation · Rakib Al-Fahad, Cloud Systems and Solutions Engineer, Intel Corporation

Abstract

Memory is rapidly becoming a dominant cost and capacity constraint in modern cloud infrastructure, driven by continued cloud-spend growth, AI-driven pressure on memory supply, and the need to support larger in-memory working sets. As a result, effective memory expansion is becoming a critical data-center priority. Linux compressed-memory mechanisms such as zswap and zram provide practical software paths for reducing swap I/O or storing compressed pages in RAM, while hardware-assisted paths such as Intel® In-Memory Analytics Accelerator (IAA) can reduce the CPU cost of compression and decompression. Building on these trends, this tutorial presents a methodology-first framework for characterizing memory compression in Linux, with emphasis on designing repeatable experiments, applying controlled memory pressure, and interpreting workload behavior across zswap, zram, and representative hardware-assisted compression paths.

The session focuses on practical, system-level evaluation of compressed-memory behavior across workloads and system configurations. It walks attendees through selecting representative workload profiles, designing static sweep and dynamic squeeze studies, collecting time-aligned telemetry, and correlating key metrics such as compression ratio, effective memory savings, CPU overhead, swap activity, throughput, average latency, and tail latency. Intel® In-Memory Analytics Accelerator (IAA) is introduced only as one example of how hardware-assisted compression can be evaluated within the same framework. Through Intel® Memory Usage Analyzer and case studies such as madvice, Redis, and other memory-intensive workloads, attendees will learn how to identify when memory compression improves effective capacity and where it can be deployed without compromising performance.

Learning Objectives

  • Understand why memory capacity and cost are becoming key constraints for cloud and data-center workloads, and why memory compression is an important evaluation target.
  • Describe a methodology-first approach for characterizing memory compression in Linux, including workload selection, baseline definition, experiment control, and result interpretation.
  • Design controlled memory-pressure experiments using cgroup v2 limits, static sweeps, and dynamic squeeze studies to expose workload sensitivity to memory reduction.
  • Collect and correlate time-aligned telemetry for compression ratio, memory savings, CPU overhead, swap activity, latency, throughput, and tail-latency behavior.
  • Compare zswap, zram, and representative hardware-assisted compression paths within the same characterization framework, using Intel® IAA only as an example implementation.
  • Use Intel® Memory Usage Analyzer and workload case studies such as madvice, Redis, and other memory-intensive workloads to identify when compression is beneficial and where deployment remains performance-safe.

Intended Audience and Prerequisites

This tutorial is intended for computer-architecture and systems researchers, graduate students, performance and capacity engineers, Linux kernel and platform developers, and cloud infrastructure practitioners interested in memory efficiency, compressed memory, and workload characterization.

Familiarity with Linux systems, basic operating-system memory management, or data-center workload performance analysis is helpful, but not required. The session introduces the required background on zswap, zram, cgroup v2 memory controls, and representative hardware-assisted compression concepts from first principles, using Intel® In-Memory Analytics Accelerator (IAA) as one example implementation. No special hardware is required to follow the methodology, and the open-source tools can be reproduced later on attendees’ own systems.

Agenda

Time Duration Topic Content
8:00 15 min Motivation and evaluation goals Memory-capacity trends, cloud cost pressure, workload impact, and the need for a repeatable compression-characterization framework
8:15 30 min Linux compression mechanisms and observability zswap, zram, reclaim behavior, MGLRU, DAMON, cgroup v2 controls, and the kernel/runtime signals needed for evaluation
8:45 25 min Methodology: controlled workload characterization Choosing representative workloads, defining baselines, applying cgroup v2 memory limits, and running static sweep and dynamic squeeze experiments
9:10 15 min Telemetry, metrics, and trade-off analysis Correlating compression ratio, effective capacity gain, CPU overhead, swap behavior, throughput, average latency, and tail latency over time
9:25 20 min Case studies and cross-path comparison Applying the methodology to madvice, Redis, zswap, zram, and one hardware-assisted compression example using Intel® IAA
9:45 15 min Decision framework, takeaways, and Q&A Determining when compression is beneficial, identifying performance-safe operating points, and adapting the methodology to new systems and workloads

Total duration: 120 minutes. Timings are indicative and can be adjusted to the final program; a short break can be inserted at the 9:10 transition if the organizers prefer.

About the Presenters

Binuraj Ravindran is a Principal Engineer in Intel’s Data Center Group specializing in system architecture, memory and resource management, AI accelerator performance modeling, and heterogeneous computing. His work spans workload characterization, scalable system design, performance optimization, and custom accelerator technologies for modern data center platforms. Contact: binuraj.ravindran@intel.com

Rakib Al-Fahad is a Cloud Systems and Solutions Engineer at Intel Corporation in Oregon, specializing in memory-usage analysis, memory compression, benchmarking, and characterization of memory-intensive data-center workloads. He contributes to the open-source Intel® Memory Usage Analyzer and helped develop the controlled memory-pressure tooling used in this tutorial, including static sweep and dynamic squeeze modes for zswap, zram, and hardware-assisted compression evaluations. Contact: rakib.al-fahad@intel.com

Materials and Resources

  • Intel® Memory Usage Analyzer (open source)
  • Kernel documentation: zswap, the IAA crypto driver, cgroup v2, and multi-generational LRU at docs.kernel.org
  • Slides and hands-on scripts will be made available to attendees through the tutorial repository.

References

  1. B. Ravindran, B. Seshasayee, K. Sridhar, R. Al-Fahad, V. Gopal, Pallavi G, and M. Chowdhury, “Memory Tiering with Intel® In-Memory Analytics Accelerator (Intel® IAA): Addressing Memory Challenges — Enhancing Data Center Efficiency with zswap and Intel IAA,” Intel Technical Paper, Doc. 846438 (PDF). Reports 2x–7x swap-in and swap-out latency improvements with Intel IAA over software compressors.
  2. “Improve Effective Memory Capacity with Efficient In-Memory Compression using zswap and Intel® IAA,” Intel Corporation — companion overview of memory-tiering with zswap and Intel IAA for effective-capacity expansion and TCO reduction in cloud and edge deployments.
  3. Intel Corporation, “Memory Tiering with Intel® In-Memory Analytics Accelerator (Intel® IAA),” Intel Technology Guide.
  4. Intel® Memory Usage Analyzer, open-source workload characterization framework.
  5. B. Ravindran et al., “Methods for Characterizing Workloads with Hardware-Accelerated Memory-Page Compression,” Tutorial, IEEE International Symposium on Workload Characterization (IISWC), 2022.
  6. K. P. Sridhar et al., “zswap compression batching with the optimized iaa_crypto driver,” Linux kernel patch series, 2025.
  7. J. Weiner et al., “TMO: Transparent Memory Offloading in Datacenters,” ASPLOS, 2022.
  8. A. Lagar-Cavilla et al., “Software-Defined Far Memory in Warehouse-Scale Computers,” ASPLOS, 2019.