《Java 并发编程·性能诊断与事故》14.3 K8s 下的并发参数基线

import java.util.concurrent.atomic.AtomicLongArray;

/** 4 个自旋线程跑 3 秒，统计相邻两次 nanoTime 的最大间隔。 */
public class SpinGap {
    public static void main(String[] args) throws Exception {
        int threads = 4;
        long deadline = System.nanoTime() + 3_000_000_000L;
        AtomicLongArray maxGap = new AtomicLongArray(threads);
        AtomicLongArray over10ms = new AtomicLongArray(threads);
        Thread[] ts = new Thread[threads];
        for (int i = 0; i < threads; i++) {
            final int id = i;
            ts[i] = Thread.ofPlatform().name("spin-" + i).start(() -> {
                long last = System.nanoTime();
                long max = 0, over = 0;
                while (System.nanoTime() < deadline) {
                    long now = System.nanoTime();
                    long gap = now - last;
                    if (gap > max) max = gap;
                    if (gap > 10_000_000L) over++;
                    last = now;
                }
                maxGap.set(id, max);
                over10ms.set(id, over);
            });
        }
        for (Thread t : ts) t.join();
        for (int i = 0; i < threads; i++) {
            System.out.print("thread=" + i);
            System.out.print(" maxGapMs=" + (maxGap.get(i) / 1_000_000));
            System.out.println(" gapsOver10ms=" + over10ms.get(i));
        }
    }
}

无限制（4 核）                  maxGapMs=6   gapsOver10ms=0
--cpus=2（4 个自旋线程）        maxGapMs=74  gapsOver10ms=29
--cpus=1                        maxGapMs=85  gapsOver10ms=30
--cpus=0.5                      maxGapMs=97  gapsOver10ms=30

--cpus=2    nr_periods 31  nr_throttled 30  throttled_usec 5968662
--cpus=0.5  nr_periods 32  nr_throttled 31  throttled_usec 10441363

apiVersion: apps/v1
kind: Deployment
metadata:
  name: libraryd-conc
  labels:
    app: libraryd-conc
spec:
  replicas: 3
  selector:
    matchLabels:
      app: libraryd-conc
  template:
    metadata:
      labels:
        app: libraryd-conc
    spec:
      terminationGracePeriodSeconds: 30
      containers:
        - name: app
          image: registry.example.com/libraryd-conc:1.0.0
          ports:
            - name: http
              containerPort: 8080
          env:
            - name: JAVA_TOOL_OPTIONS
              value: >-
                -XX:MaxRAMPercentage=70
                -XX:ActiveProcessorCount=2
                -XX:+ExitOnOutOfMemoryError
                -XX:+HeapDumpOnOutOfMemoryError
                -XX:HeapDumpPath=/var/dumps
          resources:
            requests:
              cpu: "2"
              memory: "4Gi"
            limits:
              cpu: "2"
              memory: "4Gi"
          startupProbe:
            httpGet:
              path: /healthz
              port: http
            periodSeconds: 2
            failureThreshold: 30
            timeoutSeconds: 1
          livenessProbe:
            httpGet:
              path: /healthz
              port: http
            periodSeconds: 10
            failureThreshold: 3
            timeoutSeconds: 2
          readinessProbe:
            httpGet:
              path: /readyz
              port: http
            periodSeconds: 5
            failureThreshold: 2
            timeoutSeconds: 1
          lifecycle:
            preStop:
              exec:
                command: ["sh", "-c", "sleep 5"]
          volumeMounts:
            - name: dumps
              mountPath: /var/dumps
      volumes:
        - name: dumps
          emptyDir: {}

kubectl exec deploy/libraryd-conc -- cat /sys/fs/cgroup/cpu.max
kubectl exec deploy/libraryd-conc -- cat /sys/fs/cgroup/cpu.stat
kubectl exec deploy/libraryd-conc -- cat /sys/fs/cgroup/pids.current

package com.plume.conc.ops;

/** 解析 cgroup v2 的 cpu.max 与 cpu.stat，换算成「几个核」和「节流比例」。 */
public record CpuQuota(long quotaMicros, long periodMicros) {

    public static CpuQuota parseV2(String cpuMax) {
        String[] fields = cpuMax.trim().split("\\s+");
        if (fields.length != 2) {
            throw new IllegalArgumentException("cpu.max 需要两个字段: " + cpuMax);
        }
        long period = Long.parseLong(fields[1]);
        long quota = fields[0].equals("max") ? -1 : Long.parseLong(fields[0]);
        return new CpuQuota(quota, period);
    }

    public boolean limited() {
        return quotaMicros > 0;
    }

    /** 与 HotSpot 的口径一致：向上取整，最少 1。 */
    public int visibleProcessors() {
        return limited() ? (int) Math.max(1, (quotaMicros + periodMicros - 1) / periodMicros) : -1;
    }

    public double quotaCores() {
        return limited() ? (double) quotaMicros / periodMicros : -1;
    }

    /** nr_throttled / nr_periods；持续大于 0.2 就该看线程数和 GC 了。 */
    public static double throttleRatio(long periods, long throttled) {
        return periods == 0 ? 0 : (double) throttled / periods;
    }

    public static void main(String[] args) {
        for (String s : new String[]{"200000 100000", "50000 100000", "max 100000"}) {
            CpuQuota q = CpuQuota.parseV2(s);
            System.out.printf("%-16s limited=%b quotaCores=%.2f visibleProcessors=%d%n",
                    s, q.limited(), q.quotaCores(), q.visibleProcessors());
        }
        System.out.printf("throttleRatio(31,30)=%.2f%n", throttleRatio(31, 30));
    }
}

200000 100000    limited=true  quotaCores=2.00 visibleProcessors=2
50000 100000     limited=true  quotaCores=0.50 visibleProcessors=1
max 100000       limited=false quotaCores=-1.00 visibleProcessors=-1
throttleRatio(31,30)=0.97