Project: Comparing Linux CPU Schedulers (CFS vs. EEVDF)
In this project, you are to compare two Linux CPU schedulers, the Completely Fair Scheduler (CFS) and the Earliest Eligible Virtual Deadline First (EEVDF). You will use standard scheduler benchmarks, which were used to justify Linux’s switch from CFS to EEVDF. The objectives are:
- To gain experience in measuring a system’s behavior under controlled conditions.
- To understand how a scheduling policy is chosen and how a new policy is designed, by comparing two policies that optimize different criteria.
- To connect the measurements to the algorithms discussed in the scheduling lecture.
- To document an empirical study and interpret its results, including the cases where the two policies differ and the cases where they do not.
Project Description
Linux has used the Completely Fair Scheduler (CFS) as its default scheduling policy since 2007. Starting with kernel 6.6, the kernel began transitioning to EEVDF, which is also a fair-share scheduler but adds a virtual deadline to prioritize latency-sensitive tasks (kernel documentation).
Your task is to measure both schedulers on the same workload and compare their behavior. The question you are answering is not “which scheduler is faster,” rather it is “for which workload does each scheduler make the better trade-off, and under what condition that Linux’s switch to EEVDF is justified, and why?”
Project Requirements
-
The project is open-ended. You are encouraged to design additional experiments and research questions.
-
You will set up two kernels on the same hardware (or on two machines with comparable hardware): one before 6.6 (CFS) and one 6.6 or later (EEVDF). Document how you set up the two kernels and verify the kernel versions with
uname -r. A simple method is to use two releases of a Linux distribution, for example:- Debian 12, which uses kernel 6.1 (CFS), and
- Debian 13, which uses kernel 6.12 (EEVDF).
Verify the kernel version on each machine with
uname -r, and confirm which scheduler is active from the kernel documentation (EEVDF). -
You will run at least three standard scheduler benchmarks on both kernels, with tasks pinned to a fixed core set and the machine otherwise quiet. Suggested benchmarks:
hackbench— scheduler throughput under many short-lived tasks. Distributed as part ofrt-tests: https://git.kernel.org/pub/scm/utils/rt-tests/rt-tests.gitschbench— wakeup latency and request latency. Developed by Chris Mason (then at Facebook); the maintained repository is https://github.com/masoncl/schbenchwill-it-scale(context_switch1) — context-switch throughput in thread and process modes: https://github.com/antonblanchard/will-it-scalestress-ng— scheduler pressure under a mix of workloads: https://github.com/ColinIanKing/stress-ng
You may substitute a benchmark of your own design if you justify it. For background on how these benchmarks are used by kernel developers, see the LWN survey of scheduler benchmarks (Corbet, 2017).
-
You will control for the environment: CPU frequency scaling, the number of cores, background load, and task pinning. Document each control and explain why it matters.
-
You will repeat each measurement at least 30 times and report the mean and standard deviation (or standard error) for each configuration. Plot the results with error bars. See Empirical Methodology.
-
You will compare the two schedulers using a hypothesis test appropriate for paired measurements, and report the test statistic, p-value, and effect size.
-
You shall complete a report (no more than 3 pages) documenting the following:
- What CFS and EEVDF are, and how they differ in their scheduling rule.
- The benchmarks you ran, the controls you applied, and the measurements you collected.
- The result for each benchmark, with plots and the hypothesis test.
- An interpretation of the results in terms of the scheduling criteria from the lecture (throughput, latency, turnaround time, fairness).
- A recommendation: for which workload would you choose EEVDF, and for which would you choose CFS? Justify the recommendation with your data.
Use either the ACM conference or IEEE conference proceedings template in writing the report. The ACM templates are at (ACM Conference Proceedings Template, 2026), and the IEEE templates are at (IEEE Conference Publishing Templates, 2026).
-
You shall prepare a deck of slides (no more than 6 slides, including the title and references) and make an oral presentation in class.
Special Considerations
- Scheduler benchmarks are sensitive to the environment. A machine with background load, frequency scaling, or unpinned tasks will produce noisy results. Pin the tasks and quiet the machine before measuring.
- The two kernels may have different default configurations (e.g.,
different preemption models, different tick rates). Record the
kernel version (
uname -r) and any relevant configuration (/boot/config-$(uname -r)) for both. - The benchmark results in the kernel mailing list discussion (Hegde, 2023) compare EEVDF and CFS on PowerPC hardware. Your hardware will differ. Compare the pattern of the results, not the specific numbers.
Empirical Methodology
A single run of a benchmark is not evidence. The following are required:
-
Repeat each configuration. Run each (scheduler, benchmark) configuration at least 30 times. Report the mean and the standard deviation (or the standard error of the mean).
-
Plot the distribution. A mean with error bars is acceptable; a box plot or violin plot, in addition, is more informative.
-
Use a hypothesis test for the comparison. You are comparing two paired conditions: for each benchmark, the CFS measurement and the EEVDF measurement are on the same hardware under the same conditions. Use the Wilcoxon signed-rank test or Mann-Whitney U test to test whether the two schedulers differ significantly. Report the test statistic, p-value, and effect size, and state the null and alternative hypotheses.
-
Report negative results. If a benchmark shows no significant difference between the two schedulers, report that. “No difference” is a finding, not a failure, and it is informative: it says that the two schedulers make the same trade-off for that workload.
Project Invitation
The project invitation will be distributed via CUNY’s Learning Management System, Brightspace.
Submission
Submit your work via the git repository hosted on GitHub.
- In the project repository, create a directory called
doc. - In the
docdirectory, add your project report (no more than 3 pages). - In the
docdirectory, add your presentation slides (no more than 6 slides). - Push your local repository to the remote repository on GitHub.
- Make sure you observe the deadline posted at the class website.
- Make an oral presentation in class (on a designated date).
Questions to Answer in ANSWERS.md
Each answer should be a short paragraph. Where a question asks you to show something, include the command and its output.
- Replication and Production. Can you reproduce result in the LKML mailing list discussion (Hegde, 2023), and under what condition the Linux’s switch to EEVDF is justified?
- The scheduling rule. What rule does CFS use to pick the next task? What rule does EEVDF use? What does EEVDF’s “virtual deadline” add to CFS’s “virtual runtime”?
- The kernel version. What kernel version did you run for each scheduler? How did you confirm which scheduler was active?
- The controls. What did you do to reduce noise (pinning, quiet machine, frequency scaling)? Why does each control matter?
- The hypothesis test. Which benchmark showed the largest difference between the two schedulers, and what was the p-value and effect size? Which benchmark showed no significant difference?
- The interpretation. For which workload would you choose EEVDF, and for which would you choose CFS? Justify the recommendation with your data and with the scheduling criteria from the lecture.
Further Reading
- The EEVDF scheduler documentation (EEVDF Scheduler, n.d.) — the kernel documentation, including the design rationale and the transition from CFS.
- The LKML discussion cited by the documentation (Hegde, 2023) — the benchmark comparison that justified EEVDF on PowerPC hardware.
- The CFS scheduler documentation (CFS Scheduler, n.d.) — the design of CFS, for comparison.
- The LWN survey of scheduler benchmarks
(Corbet, 2017) — background on
hackbench,schbench, and the other tools kernel developers use to evaluate scheduler patches, including whyschbenchreports p99 latencies rather than averages.
References
- EEVDF Scheduler. The Linux Kernel documentation. https://docs.kernel.org/scheduler/sched-eevdf.html
- CFS Scheduler. The Linux Kernel documentation. https://docs.kernel.org/scheduler/sched-design-CFS.html
- Hegde, S. (2023). Re: [RFC PATCH 00/10] EEVDF Scheduler. Linux Kernel Mailing List. https://lore.kernel.org/lkml/a79014e6-ea83-b316-1e12-2ae056bda6fa@linux.vnet.ibm.com/
- Corbet, J. (2017). A Survey of Scheduler Benchmarks. LWN.net. https://lwn.net/Articles/725238/
- ACM Conference Proceedings Template. (2026). https://www.acm.org/publications/proceedings-template
- IEEE Conference Publishing Templates. (2026). https://www.ieee.org/conferences/publishing/templates.html