pith. sign in

arxiv: 1906.10239 · v1 · pith:GODDJKPYnew · submitted 2019-06-24 · 💻 cs.DC · cs.OS

Container Density Improvements with Dynamic Memory Extension using NAND Flash

Jan S. Rellermeyer , Maher Amer , Richard Smutzer , Karthick Rajamani This is my paper

Pith reviewed 2026-05-25 16:43 UTC · model grok-4.3

classification 💻 cs.DC cs.OS
keywords container densitydynamic memory extensionNAND FlashDMXDockervirtual memory managementserver utilizationMemory1 system
0
0 comments X

The pith

DMX technology using NAND Flash keeps performance of critical container workloads stable even as more containers are added under constant total load.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper shows that virtual memory management becomes the bottleneck for container density on servers because many low-load containers still pressure DRAM and slow down the few high-load ones, even when their working sets would fit in main memory. DMX addresses this by transparently moving memory pages to and from NAND Flash attached through the memory bus according to a prediction model. The authors' benchmark demonstrates that in standard Docker setups, adding containers degrades critical application performance under fixed overall workload. With DMX enabled in the Memory1 system, that degradation disappears and critical performance holds steady.

Core claim

The authors describe Diablo Memory Expansion (DMX) which transparently moves memory pages between DRAM and NAND Flash attached through the memory bus using a prediction model. They present a benchmark showing that in Docker, adding additional containers adversely affects performance-critical applications even under constant overall workload. When using DMX in the Memory1 system, however, the performance of the critical workload remains stable.

What carries the argument

Diablo Memory Expansion (DMX), a prediction-model-driven mechanism that moves memory pages to and from NAND Flash via the memory bus to relieve pressure on DRAM.

If this is right

  • Performance of critical workloads stays stable when more containers are added.
  • Container density can increase without requiring additional DRAM capacity.
  • Virtual memory management ceases to be the limiting factor on server utilization in mixed-load scenarios.
  • The same overall workload can be spread across more containers while preserving response times for important applications.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same page-migration approach could be applied to other non-volatile memory types attached on the memory bus.
  • If the prediction model generalizes, similar density gains might appear in virtual-machine rather than container environments.
  • Cloud operators could reduce memory over-provisioning by relying on the extension layer for bursty or low-priority workloads.

Load-bearing premise

The prediction model can correctly identify which memory pages can be moved to NAND Flash without causing noticeable slowdowns for performance-critical containers.

What would settle it

An experiment in which critical container performance still degrades when additional containers are added despite DMX being active, or in which the model moves pages that produce measurable latency increases in high-load applications.

Figures

Figures reproduced from arXiv: 1906.10239 by Jan S. Rellermeyer, Karthick Rajamani, Maher Amer, Richard Smutzer.

Figure 1
Figure 1. Figure 1: Throughput and Latency of the AcmeAir Workload Instance [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Structure of the Memory1 DMX System and placement more e!ciently. In contrast to, e.g., the Linux Swap system which relies fully on a page fault handler and reacts the memory pressure, DMX is continually running and monitoring memory tra!c from the time an application makes memory allocation request until the process termi￾nates. Throughout that time, DMX monitors memory tra!c (without being in the memory … view at source ↗
Figure 3
Figure 3. Figure 3: Throughput and Latency of the AcmeAir Workload Instance with DMX [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Latency of the AcmeAir Workload Instance with DMX [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
read the original abstract

While containers efficiently implement the idea of operating-system-level application virtualization, they are often insufficient to increase the server utilization to a desirable level. The reason is that in practice many containerized applications experience a limited amount of load while there are few containers with a high load. In such a scenario, the virtual memory management system can become the limiting factor to container density even though the working set of active containers would fit into main memory. In this paper, we describe and evaluate a system for transparently moving memory pages in and out of DRAM and to a NAND Flash medium which is attached through the memory bus. This technique, called Diablo Memory Expansion (DMX), operates on a prediction model and is able to relieve the pressure on the memory system. We present a benchmark for container density and show that even under an overall constant workload, adding additional containers adversely affects performance-critical applications in Docker. When using the DMX technology of the Memory1 system, however, the performance of the critical workload remains stable.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 0 minor

Summary. The paper introduces Diablo Memory Expansion (DMX), a system that transparently migrates memory pages between DRAM and memory-bus-attached NAND Flash using a prediction model. It presents a container-density benchmark demonstrating that, under constant aggregate workload, adding containers in standard Docker degrades performance of critical applications due to memory pressure; the central claim is that DMX maintains stable performance for those critical workloads.

Significance. If the empirical stability result holds under rigorous validation, the work would be significant for improving server utilization in containerized environments by relieving virtual-memory bottlenecks without requiring changes to applications or the OS. The memory-bus flash attachment offers a low-latency extension path that could complement existing tiered-memory techniques.

major comments (2)
  1. [Abstract] Abstract: the assertion that 'the performance of the critical workload remains stable' with DMX supplies no quantitative metrics (throughput, latency, slowdown percentages), error bars, workload descriptions, or container counts, rendering the central empirical claim an unevaluated assertion.
  2. [Prediction model description] Prediction model description (throughout): the stability claim rests on the model correctly identifying cold pages for migration to NAND Flash, yet no precision/recall figures, false-positive migration rates, or measured slowdowns on misclassified hot pages from critical containers are reported.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments, which identify opportunities to strengthen the presentation of our empirical results. We address each major comment below and will revise the manuscript accordingly.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the assertion that 'the performance of the critical workload remains stable' with DMX supplies no quantitative metrics (throughput, latency, slowdown percentages), error bars, workload descriptions, or container counts, rendering the central empirical claim an unevaluated assertion.

    Authors: We agree that the abstract would be strengthened by the inclusion of specific quantitative metrics. In the revised manuscript we will update the abstract to report key throughput and latency values, slowdown percentages, error bars, workload descriptions, and the container counts used in the density benchmark. revision: yes

  2. Referee: [Prediction model description] Prediction model description (throughout): the stability claim rests on the model correctly identifying cold pages for migration to NAND Flash, yet no precision/recall figures, false-positive migration rates, or measured slowdowns on misclassified hot pages from critical containers are reported.

    Authors: The manuscript presents end-to-end container-density results under DMX but does not report per-model classification statistics. We will add these metrics in a revision, including precision/recall for the page-migration predictor, observed false-positive rates, and any measured slowdowns attributable to misclassified hot pages from critical containers. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical stability claim is measured outcome

full rationale

The paper presents DMX as a system that transparently migrates pages using an unspecified prediction model and then reports an empirical benchmark result: performance of critical containers remains stable under increasing container count at fixed aggregate load. No equations, fitted parameters, or derivations are described that would reduce the stability observation to a self-referential definition or to a prediction forced by the same data used to tune the model. The central claim is an experimental measurement, not a mathematical reduction, and no self-citation load-bearing steps appear in the abstract or described content.

Axiom & Free-Parameter Ledger

1 free parameters · 0 axioms · 1 invented entities

Abstract-only review; the prediction model is invoked but its internal parameters and training data are not described, so the ledger remains minimal.

free parameters (1)
  • prediction model parameters
    The abstract states that DMX operates on a prediction model whose concrete parameters are not disclosed.
invented entities (1)
  • DMX (Diablo Memory Expansion) no independent evidence
    purpose: Transparent page movement between DRAM and memory-bus NAND Flash
    Presented as the core technology enabling the density improvement.

pith-pipeline@v0.9.0 · 5708 in / 1131 out tokens · 29322 ms · 2026-05-25T16:43:03.236522+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

6 extracted references · 6 canonical work pages

  1. [1]

    [2]B￿￿￿￿, A., ￿￿￿ P￿￿, V

    [1]AcmeAir Benchmark.https://github.com/acmeair/acmeair. [2]B￿￿￿￿, A., ￿￿￿ P￿￿, V. S.SSDAlloc: hybrid SSD/RAM memory man- agement made easy. InProceedings of the 8th USENIX conference on Networked systems design and implementation(2011), USENIX Associ- ation, pp. 211–224. [3]B￿￿￿￿￿￿, I., C￿￿￿￿￿, P., C￿￿￿￿, K., C￿￿￿￿, P., F￿￿￿, S., I￿￿￿￿￿￿￿, V., M￿￿￿￿￿￿￿, ...

    work page 2011

  2. [2]

    T., ￿￿￿ K￿￿￿￿￿, G.The Apache HTTP Server Project

    [9]F￿￿￿￿￿￿￿, R. T., ￿￿￿ K￿￿￿￿￿, G.The Apache HTTP Server Project. IEEE Internet Computing 1, 4 (Jul 1997), 88–90. [10]H￿￿￿￿￿, E. H.Apache JMeter: A practical beginner’s guide to auto- mated testing and performance measurement for your websites. Packt Publishing Ltd,

    work page 1997

  3. [3]

    J.Experiments on Page Size, Program Access Patterns, and Virtual Memory Performance.IBM Journal of Research and Devel- opment 16, 1 (Jan 1972), 58–66

    [11]H￿￿￿￿￿￿￿, D. J.Experiments on Page Size, Program Access Patterns, and Virtual Memory Performance.IBM Journal of Research and Devel- opment 16, 1 (Jan 1972), 58–66. [12]J￿￿￿, D., K￿￿, J.￿￿., P￿￿￿, S.￿￿., K￿￿￿, J.￿U., ￿￿￿ L￿￿, J.Fass: A ￿ash- aware swap system. InProc. of International Workshop on Software Support for Portable Storage (IWSSPS)(2005). [1...

    work page 1972

  4. [4]

    O’Reilly Media, Inc

    ICCSA’08. International Conference on(2008), IEEE, pp. 151–156. [14]Memory1. https://www.supermicro.com.tw/solutions/datasheet_ Diablo_Memory1.pdf. [15]M￿￿￿￿￿, D.Docker: Lightweight Linux Containers for Consistent Development and Deployment.Linux J. 2014, 239 (Mar. 2014). [16]M￿￿￿￿￿￿, M., C￿￿￿, F., L￿￿, T., ￿￿￿ A￿￿￿￿, J. B.Di￿erentiated storage services. ...

    work page 2008

  5. [5]

    M.FlashVM: Virtual Memory Management on Flash

    [21]S￿￿￿￿￿, M., ￿￿￿ S￿￿￿￿, M. M.FlashVM: Virtual Memory Management on Flash. InUSENIX Annual Technical Conference(2010). [22]S￿￿￿￿￿￿, S., P￿￿￿￿, H., F￿￿￿￿￿￿￿￿￿, M. E., B￿￿￿￿￿, A., ￿￿￿ P￿￿￿￿￿￿￿, L.Container-based operating system virtualization: a scalable, high- performance alternative to hypervisors. InACM SIGOPS Operating Systems Review(2007), vol. 41, ...

    work page 2010

  6. [6]

    ISPASS 2005.(March 2005), pp. 248–257. [25]X￿￿￿￿￿, M. G., D￿ O￿￿￿￿￿￿￿, I. C., R￿￿￿￿, F. D., D￿￿ P￿￿￿￿￿, R. D., M￿￿￿￿￿￿￿￿, K. J., ￿￿￿ D￿ R￿￿￿, C. A.A performance isolation analysis of disk-intensive workloads on container-based clouds. InParallel, Distributed and Network-Based Processing (PDP), 2015 23rd Euromicro International Conference on(2015), IEEE, p...

    work page 2005