Skip to content

README.md

Latest commit

 

History

History
113 lines (77 loc) · 5.9 KB

README.md

File metadata and controls

113 lines (77 loc) · 5.9 KB

Artifact Evaluation for "Xkernel: Rethinking Performance Tunability of Operating System Kernels" (OSDI '26)

1. Overview

1.1 Artifact Goals

In this paper, we make the following claims:

  • Xkernel enables tuning consts for adapting to hardware devices and workload patterns. (i.e., BLK_MAX_REQUEST_COUNT in Figure 1)
  • Xkernel enables tuning consts for balancing cost-benefit tradeoffs. (i.e., MAX_SOFT_IRQ_RESTART in Figure 9)
  • Xkernel enables tuning consts for controlling kernel internal behavior. (i.e., SHRINK_BATCH in Figure 10)
  • Xkernel enables tuning consts with kernel and hardware observability. (i.e., NR_MAX_BATCHED_MIGRATION in Figure 11)
  • Xkernel enables collective tuning of interdependent consts. (i.e., HYSTART_DELAY_[MAX,MIN,factor] in Figure 12)
  • Xkernel introduces negligible overhead at runtime. (i.e., Figure 16)
  • Xkernel has lower transition latency than Linux KLP. (i.e., Figure 18)

We also provide a dataset characterization of perf-consts in the Linux kernel (Figures 13-14) and ad-hoc design drill-down measurements for understanding the Xkernel deployment pipeline (Figure 17) and global convergence overhead (Figures 19-20).

This artifact evaluation package reproduces all figures from the paper. Each figure corresponds to a self-contained experiment with detailed instructions in its README.md (under ae/Figure*/). Figure1/9/10/11/12/16/18 needs to be reproduced by running the corresponding experiment, while Figure13-14/17/19-20 (which are dataset-related or ad-hoc design drill-down) are plot-only and can be generated from the provided raw data.

1.2 Artifact Structure

The artifact evaluation requires 5 CloudLab machines provisioned via 3 CloudLab profiles (under ae/cloudlab-profiles/).

Profile Nodes Hardware Cluster Used By
c220g5.xml 1 c220g5 Wisconsin Figure 1(a), 10, 11
c6620.xml 2 c6620 Clemson Figures 1(b), 16, 18
xl170.xml 2 xl170 Utah Figures 9, 12

Each experiment directory (ae/Figure*/) follows a common three-script structure:

  1. install_*.sh — Installs environment dependencies (benchmarks, libraries, etc.).
  2. run.sh — Runs the experiment and saves raw data to results/.
  3. plot/plot.py — Generates the corresponding figure from the paper.

Refer to the README.md inside each Figure*/ directory for detailed instructions.

Note: run.shis designed to be idempotent (safe to re-run), though rare corner cases (e.g., unexpected interruption) may require manual cleanup before retrying.

1.3 Artifact Time

  • Install Xkernel in parallel on three machines (Section 2.3) — ⏱ 20-30 min

  • Run experiments in parallel

    Machine Figures Est. Time
    c220g5 1(a) → 10 → 11 ⏱ 5 min + 20 min + 20 min
    c6620 1(b) → 16 → 18 ⏱ 5 min + 10 min + 40min
    xl170 9 → 12 ⏱ 20 min + 10 min
    Your laptop 13-14, 17, 19-20 ⏱ 5 min

If you have any questions, feel free to contact us via email or HotCRP.

2. Setting up CloudLab Machines

If you are a first-time CloudLab user, please read CloudLab For Artifact Evaluation for an overview of the process.

If you do not already have a CloudLab account, please apply for one here and ask the OSDI AEC chair to add you to the AEC project. Please let us know if you have trouble accessing CloudLab — we can help set up experiments and give you access.

2.1. Reserve Nodes

Machines may not always be available. To guarantee availability, click Experiments → Reserve Nodes from the CloudLab dashboard. Select the appropriate cluster and hardware type (see table above), specify the number of nodes and your desired time window, then submit the request. See Resource Reservation for details.

2.2. Create Experiments Using Profiles

We provide three CloudLab profile XML files under ae/cloudlab-profiles/.

To use a profile:

  1. Go to Experiments → Create Experiment Profile.
  2. Click Upload and select the corresponding XML file (e.g., xl170.xml).
  3. Give the profile a name and click Create.
  4. Click Instantiate on the profile page, then keep clicking Next / Finish.
  5. Wait for Status to become Ready on the experiment page.
  6. Access each node via ssh using the hostname shown on the experiment page.

2.3. Install Xkernel

Please install Xkernel on three machines--c220g5, c6620 with IP 192.168.100.1, and xl170 with IP 192.168.6.1 before running the experiments.

Note: There is no need to install Xkernel on all c6620 and xl170 machines.

git clone https://github.com/zhongjiechen/Xkernel.git
cd Xkernel
./xkernel-tool setup # about 20-30minutes, depending on the machine.
export KERNEL_DIR=~/linux-6.8.0

2.4. Kernel Version Policy

Most experiments in this artifact use the default Linux 6.8 environment:

export KERNEL_DIR=~/linux-6.8.0

Figure 10 and Figure 11 are the only exceptions. They should be reproduced on Linux 6.14.8-061408-generic, with Xkernel's code generation pointed at the matching Linux 6.14.8 source tree:

export KERNEL_DIR=~/linux-6.14.8-061408-generic

This version switch is local to Figure 10 and Figure 11. We recommend running the Linux 6.8 figures first, then rebooting the c220g5 machine into Linux 6.14.8-061408-generic and running Figure 10 and Figure 11. If you return to the other experiments afterward, boot back into the Linux 6.8 environment and reset KERNEL_DIR=~/linux-6.8.0.

The Figure 10 and Figure 11 README files include the exact commands for installing Linux 6.14.8-061408-generic and preparing the matching source tree.