How does the Linux Completely Fair Scheduler (CFS) use vruntime to enforce proportional CPU sharing?

The Completely Fair Scheduler (CFS) does not use traditional fixed timeslices. Instead, it tracks each task's virtual runtime (vruntime), which represents the normalized amount of CPU execution time a task has consumed. When a task runs for physical duration dt, its vruntime increases by dt * (NICE_0_LOAD / task_weight), where task_weight is derived from the nice level or cgroup cpu.weight. Tasks with higher weight (higher priority) advance their vruntime slower than lower-weight tasks. CFS indexes all runnable tasks in a time-ordered Red-Black tree; the task with the smallest vruntime (leftmost node) is always selected for execution next. This guarantees that all runnable tasks receive CPU time strictly proportional to their assigned weights over time.

Why do multithreaded containers experience severe CFS throttling even when total average CPU utilization is low?

In cgroups v2, cpu.max sets a quota over a fixed period (by default, period is 100,000 microseconds / 100ms). For example, a quota of 200,000us allows 2 cores of compute per 100ms window. However, quota consumption is aggregated across ALL concurrent threads in the cgroup. If a container with a 2-core limit spawns 16 threads that simultaneously execute work during an incoming web request burst, those 16 threads consume 160,000us of CPU time in just 10ms (10ms * 16 threads). The container exhausts its 200,000us quota in 12.5ms, and the Linux kernel forcibly unschedules and freezes all 16 threads for the remaining 87.5ms of the period! Even though average CPU utilization over the second looks low (only 20%), the application suffers catastrophic P99 tail latency spikes and timeouts.

What is cpu.max.burst in Linux 5.14+ and how does it prevent micro-throttling in Kubernetes?

Prior to Linux 5.14, any unused CPU quota at the end of a 100ms CFS period was completely discarded. In Linux 5.14+, the kernel introduced the CPU burst feature (controlled by cpu.max.burst in cgroups v2). If a container uses less than its allocated quota in period N, the surplus quota is rolled over into a burst buffer (up to a configurable maximum). When a burst of traffic arrives in period N+1, the container can borrow accumulated buffer quota to complete bursty requests without being throttled, even if instantaneous thread concurrency exceeds the baseline limit. This eliminates false-positive latency spikes without allowing containers to sustain elevated long-term CPU consumption.

How do Kubernetes cpu requests and limits map to cgroups v2 cpu.weight and cpu.max?

Kubernetes maps CPU requests and limits to distinct cgroup controllers. A CPU request (e.g. 500m) is converted to cgroups v2 cpu.weight: the formula scales the request to a 1-10000 range (weight = 1 + ((requests - 2) * 9999) / 262142), ensuring proportional share when host CPU cores are 100% saturated. Conversely, a CPU limit (e.g. 2000m) maps directly to cgroups v2 cpu.max: "200000 100000" (quota of 200ms per 100ms period). Because hard limits cause aggressive CFS throttling under bursty loads, many production engineering teams (such as at Uber, Netflix, and Datadog) set CPU requests for autoscaling and bin-packing but omit CPU limits entirely, relying on cpu.weight and CFS fair-share mechanics instead.

Linux CFS Bandwidth & CPU Quota Studio: cgroups v2, cpu.max, vruntime & Throttling

Verified Specification | Standardized Formulas | Instant Precision

Secure & Private (Zero Data Retention) Free Access • No Sign-Up

Linux CFS Bandwidth & CPU Quota Studio

Architect container CPU limits and scheduling: simulate Completely Fair Scheduler vruntime progression, model cgroups v2 cpu.max quota exhaustion, and analyze multithreaded throttling.

cgroups v2 Active Throttling: 0%

2.0 Cores

Allocated CPU Bandwidth (Quota / Period)

100.0 ms

Quota Exhaustion Point in 100ms Period

0.0%

Period Throttling Percentage

1.0x (Normal)

P99 Latency Degradation Factor

1. cgroups v2 CPU Bandwidth & Workload Parameters

cpu.max Limit (Cores)

Concurrent Worker Threads

Workload Profile

cpu.max.burst Feature (Linux 5.14+)

2. 100ms CFS Period Microsecond Timeline & Quota Depletion

CFS Period Progress (0 to 100,000 µs)

■ CPU Executing (Running) ■ Kernel Throttled (Frozen) ■ Idle Capacity

Inspecting period consumption...

CFS Runqueue (cfs_rq) vruntime Balance

3. Production Container & OS Configuration

// Generated config output

Linux CFS Bandwidth & CPU Quota Studio

1. cgroups v2 CPU Bandwidth & Workload Parameters

2. 100ms CFS Period Microsecond Timeline & Quota Depletion

CFS Runqueue (cfs_rq) vruntime Balance

3. Production Container & OS Configuration

🔍 Companion Workbenches & Verification Tools

Linux Landlock LSM ABI v4 Network & Path Sandboxing Studio

URL Parser & Query String Parameter Inspector

Prometheus, PromQL & TSDB Storage Engine Architecture Studio

JavaScript Minifier & Code Compressor

Why Smart Users Are Switching to This Free Toolkit

The Browser Feature Most People Never Use

Stop Paying for Software You Can Get Free

File Conversion Tricks That Save Hours Every Week

This Free Tool Replaced 5 Paid Apps for Our Team

What Your Browser Can Do Without Any Extensions