Description
The scheduler is at the core of Linux performance. With different topologies and workloads, giving the user the best experience possible is challenging, from low latency to high throughput and from small power-constrained devices to HPC.
The following accomplishments have been made as a result of last year’s micro-conference:
- Progress on proxy execution
https://lore.kernel.org/lkml/20240224001153.2584030-1-jstultz@google.com/ - Progress on system pressure
https://lore.kernel.org/lkml/170073688055.398.12687414937207369825.tip-bot2@tip-bot2/
https://lore.kernel.org/lkml/20240220145947.1107937-1-vincent.guittot@linaro.org/ - Progress in the DL server
- The EEVDF scheduler and improvements in latency nice
- Progress on adding tracepoints for IPI
Ideas of topics to be discussed include (but are not limited to):
- Improve responsiveness for CFS tasks
- The improvements on the EEVDF scheduler proposal
- Impact of new topology on CFS, including hybrid or heterogeneous system
- Taking into account task profile with IPCC or uclamp
- Locking improvements – e.g., proxy execution
- Improvements on SCHED_DEADLINE
- Tooling for debugging scheduling
It is fine if you have a new topic not on the list. People are encouraged to submit any topic related to real-time and scheduling.
The goal is to discuss open problems, preferably with patch set submissions already in discussion on LKML. The presentations are concise, and the central portion of the time should be given to the debate – thus, the importance of having an open and relevant problem with people in the community engaged in the solution.
-
Tobias Huschle18/09/2024, 15:00
In virtualized environments, information about the underlying physical CPU topology is usually hidden from the guest systems.
Go to contribution page
This talk will discuss challenges in scheduling virtual CPUs and how passing topology insights to the guests can be utilized to allow the guests to cooperate and gain performance benefits.
As an example, the s390 architecture will be used to demonstrate how smart... -
Shrikanth Hegde18/09/2024, 15:22
CPU capacity is a software construct to reflect underlying physical CPU capacity. Load balancer uses the CPU capacity to choose an optimal CPU for performance and energy efficiency. CPU Capacity can be affected by frequency, higher level sched classes, guest preemption etc. Steal time is an indicator of guest preemption by the host hypervisor. Current Linux scheduler, updates the CPU capacity...
Go to contribution page -
Mr Qais Yousef (Google)18/09/2024, 15:44
One of the bottlenecks to making progress in scheduler is understanding 'what is the problem?'
Reporters who don't understand the scheduler can't provide useful info to help root cause why they see a problem.
Developers, seasoned or new, can trip over many details and corner cases that might make what appears to be a bug to be actually a feature that is just not well understood by the...
Go to contribution page -
Xi Wang18/09/2024, 16:07
Throttling-like mechanisms such as CFS bandwidth control, extremely biased cgroup CPU shares and CPU masks can create quasi priorities among CFS tasks, and we can get priority inversion without explicit priority. We had such a problem caused by deep CPU throttling with CFS bandwidth control and it was causing application timeouts and down time.
To solve this problem we created a priority...
Go to contribution page -
Vincent Guittot (Linaro)18/09/2024, 17:00
The Energy Aware Scheduler has been developed with the assumption that tasks wake up regularly and often enough to keep their placement energy efficient and responsive. This assumption tend to be less and less true because of constraints like capping the performance of the system and the needs for still shorter response time. New mechanisms outside the wakeup path needs to be enabled to solve...
Go to contribution page -
Pintu Kumar18/09/2024, 17:22
In any system the overall system congestion behavior mainly revolve around CPU work-load, memory-pressure and IO-wait.
Go to contribution page
The Pressure Stall Information (PSI), introduced to monitor resource contention by tracking CPU, memory, and I/O pressure, provides real-time insights into system performance bottlenecks.
But the problem is, it just gives the overall average load value in the system during... -
Mr Qais Yousef (Google)18/09/2024, 17:44
Power management features like DVFS introduces Time Dilation effect where progress of time slows down the lower the frequency from the task's perspective.
Combined with Heterogeneous systems (HMP) this Time Dilation become more extreme on the smaller cores. Especially on Arm mobile SoCs where the little cores are too small on many SoCs.
This manifests as big delays in task's rampup...
Go to contribution page -
John Stultz (Google)18/09/2024, 18:05
At OSPM we had a number of discussions around the need for QoS APIs for applications to hint their needs for latency and throughput for SCHED_NORMAL/FAIR tasks, as opposed to the typical global tuning of scheduler knobs.
Folks seemed generally supportive of adding some sort of hinting API. However, while any hinting API will be coy and resistant to making any hard promises to userland...
Go to contribution page
