OMAP4 pandaboard上CPU hotplug进程迁移与负载均衡测试

尝试把pandaboard images升级为 11.12 Linaro release，SMP热插拔测试了一下。第一组测试评估在高负载情况下热插拔CPU进程迁移的情况，希望是拔出和插入CPU1后，各进程仍然活跃，测试case如下：

int main(int argc, char *argv[]) { struct timeval last_tv, cur_tv; int pid; char p_proc[50], on = '1', off = '0'; int fd = open("/sys/devices/system/cpu/cpu1/online", O_RDWR); if (fd < 0) exit (0); sprintf(p_proc, "/proc/%d", getpid()); gettimeofday (&last_tv , NULL); /* fork to 16 processes with high cpu usage */ pid = fork(); if (pid > 0) { printf("before unplugging cpu1\n"); sleep(5); /* unplug cpu1 */ write(fd, &off, 1); printf("after unplugging cpu1, we want to see all processes are still alive\n"); sleep(5); /* re-plug cpu1 */ printf("after re-plugging cpu1\n"); write(fd, &on, 1); sleep(5); } else { /* lauch multi-processes with 100% cpu usage */ fork(); fork(); fork(); fork(); while(1) { gettimeofday (&cur_tv , NULL); if (cur_tv.tv_sec - last_tv.tv_sec == 4) { printf("pid: %d\n", getpid()); last_tv = cur_tv; } /* if parent process exit, sub-process exit */ if (access(p_proc, 0) < 0) _exit(0); } } }
Fork出16个CPU利用率100%的进程，在热插拔CPU的情况下，这些16个进程保持活跃，证明热插拔的情况下，进程可迁移：

root@android:/ # smp-unplug-process-alive before unplugging cpu1 pid: 582 pid: 583 pid: 576 pid: 579 pid: 574 pid: 587 pid: 580 pid: 586 pid: 581 pid: 588 pid: 575 pid: 573 pid: 577 pid: 584 pid: 578 pid: 585 [ 68.295593] CPU1: shutdown after unplugging cpu1, we want to see all processes are still alive pid: 587 pid: 581 pid: 578 pid: 576 pid: 582 pid: 585 pid: 577 pid: 574 pid: 573 pid: 580 pid: 586 pid: 579 pid: 583 pid: 575 pid: 584 pid: 588 after re-plugging cpu1 [ 73.520935] CPU1: Booted secondary processor [ 73.528350] Switched to NOHz mode on CPU #1 pid: 588 pid: 581 pid: 582 pid: 576 pid: 586 pid: 585 pid: 579 pid: 584 pid: 587 pid: 573 pid: 578 pid: 580 pid: 577 pid: 583 pid: 575 pid: 574 …
第二组测试评估2个processes并行计算下，CPU hotplug后对时间的影响，以彻底验证拔掉CPU1后只有CPU0在运算，重新插入后2个开始一起运算，测试case如下：

int main(int argc, char *argv[]) { volatile int i; int pid = fork(); if (pid > 0) { for (i=0;i<90000000;i++); wait(0); } else { for (i=0;i<90000000;i++); } }

在不unplug cpu1的情况下：

root@android:/ # time /system/bin/smp-time 0m0.73s real 0m1.42s user 0m0.03s system root@android:/ # time /system/bin/smp-time 0m0.72s real 0m1.41s user 0m0.01s system root@android:/ # time /system/bin/smp-time 0m0.72s real 0m1.43s user 0m0.00s system
实际运行时间0.73秒，在2个CPU上累积耗费了1.42秒, 各自贡献1/2力量。
现在我们拔出cpu1，再运行：

echo 0 > /sys/devices/system/cpu/cpu1/online [ 234.565521] CPU1: shutdown root@android:/ # time /system/bin/smp-time 0m1.45s real 0m1.38s user 0m0.06s system root@android:/ # time /system/bin/smp-time 0m1.45s real 0m1.44s user 0m0.00s system root@android:/ # time /system/bin/smp-time 0m1.45s real 0m1.45s user 0m0.00s system
实际运行耗时1.45秒，等于user耗时，因为现在只有1个CPU0顺序执行2个进程。
我们把CPU1插入，再执行：

echo 1 > /sys/devices/system/cpu/cpu1/online [ 364.515075] CPU1: Booted secondary processor root@android:/ # time /system/bin/smp-time 0m0.72s real 0m1.42s user 0m0.00s system root@android:/ # time /system/bin/smp-time 0m0.72s real 0m1.43s user 0m0.00s system
真实时间又变成user时间1/2，证明CPU1活跃，开始与CPU0并行运算。
结论是Linux kernel有较好的CPU热插拔支持，在热插拔情况下，进程能够顺利迁移与再次负载均衡。