---------- Forwarded message ---------
From: "Don Harbin (Linux Fundamentals)" <linux_fundamentals(a)hub.corecollective.dev>
Date: Wednesday, 22 April 2026 at 11:21:17 UTC-7
Subject: Fwd: [Linux Fundamentals] Minutes: 2026-04-21 CoreCollective Linux Fundamentals WG
To: CoreCollective Linux Fundamentals
<linux-fundamentals(a)corecollective.dev>
Hi all,
Please find the minutes from yesterday's tech meeting below.
Vincent's slides are attached. Please let us know your feedback / updates.
Please let us know if you spot any errors or important omissions.
Thanks & regards,
Don Harbin | *Community Director*
======== Minutes ==========
Please review Gemini's notes to ensure they are accurate
*Participants*: Linaro, Suse, Google, AMD
*Summary*
Working group members reviewed scheduling interface proposals and discussed
funding processes for collaborative Linux kernel projects.
*Linux Scheduler Proposal Review*Participants evaluated the User Assisted
Scheduling Interface proposal to optimize application-scheduler
interactions. The group decided to utilize a netlink interface rather than
direct library linking to avoid feedback loops.
*Performance and Fairness Optimization*Development efforts focus on thread
tagging and Quality of Service levels to improve system responsiveness.
Testing indicates reduced latency for sensitive tasks by using Earliest
Virtual Deadline First scheduling.
*Funding and Group Processes*Leadership outlined procedures for proposing
technical initiatives to the steering committee for financial support.
Collaboration strategies address managing cross-timezone participation and
gathering industry feedback for commercial viability.
*Suggested next steps*
[Lina Iyer, Don Harbin] Distribute Materials: Share Vincents presentation
slides and meeting notes with the entire group. *DONE*[*The group*] Gather
Feedback: Share scheduling feature details with internal teams; solicit
feedback on potential use cases and necessary changes.[Lorenzo] Prepare
Presentation: Present information regarding GIC v5 development in a future
meeting.[Don Harbin] Send Follow-up Note: Send an email soliciting ideas or
thoughts on high-priority kernel gaps to the mailing list.[Lina Iyer]
Prepare Agenda Topics: Send out advanced discussion topics before the next
meeting to facilitate participation.
Details
-
Introduction and Working Group Goals: The meeting served as the second
gathering of the Linux Fundamental Working Group, with Lina Iyer welcoming
attendees and recapping that the group focuses on collaborative areas
within Linux, including the Linux kernel, kernel CI testing, and
performance improvements. The primary goal is to find collaborative
initiatives that can be funded, allowing participants to pool their efforts
on one or more projects simultaneously.
-
Presentation on Linux Scheduler and User Space Interaction: Vincent
Guittot presented work done by the Linaro team regarding the Linux
scheduler and user space. The presentation focused on improving the
interaction between user space applications and the scheduler to bridge the
gap between application expectations and scheduler details, which has been
a topic of discussion at conferences like LPC.
-
User Assisted Scheduling Interface (SKS) Proposal: A proposal called the
User Assisted Scheduling Interface (SKS) was presented, which aims to
improve the application-scheduler interaction . The initial decision was
to avoid providing an API or library linked directly to an application,
primarily because this method creates a long feedback loop for interface
updates and deployments. Instead, the approach uses a netlink interface to
track thread creation and destruction, combined with per-application
description files to define expected thread behavior.
-
Thread Tagging and Quality of Service (QoS) Levels: The new user space
interface is based on tagging threads according to their primary function,
using types such as user interactive, user initiated, utility, background,
and default. These QoS tags are then translated into specific scheduler
attributes, like scheduling policy, nice priority, and EVDF duration, to
manage system responsiveness by lowering the noise from undescribed or
background tasks.
-
Initial State and Default Behavior of Threads: Currently, the thread
tagging is static upon creation, though this is not a fixed configuration
for future versions. By default, any task that is not explicitly
described will be configured as a batch task with a low nice priority and a
long EVDF slice duration to limit their running speed and lower overall
system noise. The goal is to ensure that tasks tagged as user
interactive receive higher CPU bandwidth and run first upon waking up,
before other lower-priority activities.
-
Implementation Approach Using Netlink: Netlink was chosen as the
interface method because it allows a daemon to monitor every thread
creation and destruction event to apply the appropriate configuration based
on a configuration file. This approach, which currently uses existing C
interfaces like `set attributes`, avoids the need for new application
interfaces or linking new libraries, making testing easier and allowing for
flexibility in configuration across different systems like laptops, mobile
phones, and servers.
-
Scheduler Fairness and Latency Management: To manage latency without
compromising fairness, efforts are focused on fair task ordering by
leveraging EVDF scheduling. This ensures that short tasks, having a shorter
EVDF slice duration, will run first as long as they do not exceed their CPU
share, which has been shown to significantly reduce scheduling latency for
latency-sensitive tasks. Initial tests using cyclic test and bench
showed that setting background tasks to a 20 millisecond slice and cyclic
test to an 8 millisecond slice can reduce the cyclic test scheduling
latency down to 1 millisecond for 99.9% of cases.
-
Improvements to Scheduling Push Balance and Priority Inheritance:
Ongoing work also includes improving the scheduling push balance to select
per-task CPU more often, using the same consistent algorithm for both
wake-up and periodic load balancing. Additionally, there is a need to
fix priority inheritance across all scheduling classes, including SKS, to
ensure higher-priority tasks can make progress by enforcing priority
inheritance in user space by default.
-
Areas for Participation and Future Development: Vincent Guittot invited
feedback on the SKS interface, reporting test results, and developing
missing features. Currently, the management of grouping and window
management, as well as the extension of priority inheritance to
non-blocking locks, requires more development.
-
Solicitation for Industry Feedback and Business Case Discussion: Lina
Iyer emphasized that the working group's intent is to gather feedback from
various industries and domains to ensure that new features, such as the
SKS, are commercially viable for different use cases. The forum is ideal
for discussing the business-related viewpoint of these use cases, such as
how to handle frequently changing priorities or QoS levels in applications
.
-
Upcoming Topic and Initiative Funding: The next meeting will potentially
feature a discussion on GEEK v5, an initiative financially supported by
ARM. The group also discussed the process for initiating new work streams,
where consensus on a technical challenge could lead to an initiative being
proposed to the steering committee for potential funding.
-
Process for Proposing New Initiatives and Communication: Don Harbin
explained that new working groups or initiatives can be proposed by sending
an outline of the technical challenge, proposed leadership, and interested
companies to the inquiries mailing list. Participants were also
encouraged to reach out to Don Harbin or Lina Iyer directly with ideas or
high-priority technology gaps for the kernel.
Timezone Consideration and Collaboration Strategy: Lina Iyer noted the
challenge of scheduling the meeting across multiple time zones and invited
suggestions on how to maintain synergy without fracturing the discussion
across different meeting times. Randy Linnell reinforced that engagement
and involvement are the way for participants to get value from the group,
which is intended to be highly collaborative.