- Preemption Context Switches度量的是操作系统任务调度器将处理器中的一个正在运行的线程切换为另一个更高优先级的线程的次数,即发生抢占的次数。
- Synchronization context switches度量的是由于显式调用线程同步API而发生线程切换的次数,如给多线程共享的变量加锁,多线程共同去修改,有些线程要阻塞在lock,直至占用锁的线程释放lock,这个度量反映的是线程间竞争的程度。
下面的实验来自VTune,旨在探究Preemption Context Switches的来源。
实验一:多线程无锁保护
speedup-example-no-mutex.cpp
#include <pthread.h>
#include <stdio.h>
#include <stdlib.h>
#include <errno.h>
#include <assert.h>
#define N 4
#define M 30000
int nwait = 0;
volatile long long sum;
long loops = 6e3;
void set_affinity(int core_id) {
cpu_set_t cpuset;
CPU_ZERO(&cpuset);
CPU_SET(core_id, &cpuset);
assert(pthread_setaffinity_np(pthread_self(), sizeof(cpu_set_t), &cpuset) == 0);
}
void* thread_func(void *arg) {
set_affinity((int)(long)arg);
for (int j = 0; j < M; j++) {
nwait++;
for (long i = 0; i < loops; i++)
sum += i;
for (long i = 0; i < loops; i++)
sum += i*i*i*i*i*i;
}
}
int main(int argc, char *argv[]) {
set_affinity(23);
pthread_t th[N];
int ret;
for(unsigned i=0; i<N; ++i) {
ret = pthread_create(&th[i], NULL, thread_func, (void*)i);
assert(!ret && "pthread_create() failed!");
}
for(unsigned i=0; i<N; ++i)
pthread_join(th[i], NULL);
exit(0);
}
VTune现象:
Preemption Context Switches由两部分组成:clone和Unknown stack frame(s)。
- 后者的Preemption稳定在5:在这个程序中,共有5个线程在运行,VTune显示每个线程各占1,所以后者的Preemption才稳定在5上。为了验证,我们让N等于8,结果是每个线程各占1,Unknown stack frame(s)处的Preemption稳定在9。
- clone处的Preemption不是一个确定的数,有可能是6、7、8等。
通过上图可以发现clone处的Preemption都分布在四个子线程中。下面再来一组:
通过比较上面三幅图,我们发现四个子线程所占的Preemption数并不总是均等。为了验证,我们让N等于8,结果如下:
果然clone处的Preemption并不是由子线程均分。不过随着线程数增加,clone处Preemption的增加幅度要大于Unknown stack frame(s)处。
通过上面的现象,我们尝试做出结论:
由于没有锁,所以线程间是独立的,我们单独分析一个线程中Preemption Context Switches的来源即可(其实这种假设是有问题的,因为我们上面提到随着线程数增加,Preemption并没有线性增加,如果各线程间相互独立,理应是线性增加的,不过我们先从简单情况入手)。我们尝试逐步减少子线程执行任务的办法:
void* thread_func(void *arg) { set_affinity((int)(long)arg); for (int j = 0; j < M; j++) { for (long i = 0; i < loops; i++) sum += i*i*i*i*i*i; } }
void* thread_func(void *arg) { set_affinity((int)(long)arg); for (int j = 0; j < M; j++) { for (long i = 0; i < loops; i++) sum += i; } }
void* thread_func(void *arg) { set_affinity((int)(long)arg); for (int j = 0; j < M; j++) { nwait++; } }
无clone处的Preemption Context Switches
通过上面我们就断定当子线程计算任务变轻时,clone处的Preemption会变少,这是武断的。因为如下:
void* thread_func(void *arg) { set_affinity((int)(long)arg); for (int j = 0; j < M; j++) { for (long i = 0; i < loops; i++) sum += i; for (long i = 0; i < loops; i++) sum += i; } }
这个子线程的计算任务要比上面三个中的第一个要轻,但它的Preemption数却要多,所以我初步猜想是第二层for循环的个数决定了clone处的Preemption数,于是做以下验证:
void* thread_func(void *arg) { set_affinity((int)(long)arg); for (int j = 0; j < M; j++) { for (long i = 0; i < loops; i++) sum += i; for (long i = 0; i < loops; i++) sum += i; for (long i = 0; i < loops; i++) sum += i; } }
void* thread_func(void *arg) { set_affinity((int)(long)arg); for (int j = 0; j < M; j++) { for (long i = 0; i < loops; i++) sum += i; for (long i = 0; i < loops; i++) sum += i; for (long i = 0; i < loops; i++) sum += i; for (long i = 0; i < loops; i++) sum += i; } }
确实是随着for循环的增多,clone处的Preemption在增多,但以此下结论还是不妥,合理的验证还应有以下工作:
void* thread_func(void *arg) { set_affinity((int)(long)arg); for (int j = 0; j < M; j++) { for (long i = 0; i < loops; i++) { sum += i; sum += i; sum += i; sum += i; } } }
奇怪明明这个子线程的工作量和上面验证中的第二个一样,而且它只有一个for,但clone处的Preemption却更多,于是继续做验证:
void* thread_func(void *arg) { set_affinity((int)(long)arg); for (int j = 0; j < M; j++) { for (long i = 0; i < loops; i++) { sum += i; sum += i; sum += i; sum += i; sum += i; sum += i; sum += i; } } }
最终结论:
也就是说随着第二层for的个数增加,clone处Preemption在增加;如果第二层只有一个for,那么随着这个for中的子句(上面的实验只能说明本例中出现的子句sum+=i有这种情况)的增多,clone处的Preemption在增加。
分析:
如果说这是结论,那为什么?子线程在执行时,频繁被更高优先级的进程给抢占,可能是时间,运行时间,当子线程运行时间长时,系统中更高优先级的进程抢占它的情况更多。果然,我们重新执行上述那些验证程序,发现——clone处Preemption多的程序,它的运行时间越长。
void* thread_func(void *arg) { set_affinity((int)(long)arg); for (int j = 0; j < M; j++) { for (long i = 0; i < loops; i++) sum += i; for (long i = 0; i < loops; i++) sum += i; for (long i = 0; i < loops; i++) sum += i; for (long i = 0; i < loops; i++) sum += i; } }
void* thread_func(void *arg) { set_affinity((int)(long)arg); for (int j = 0; j < M; j++) { for (long i = 0; i < loops; i++) { sum += i; sum += i; sum += i; sum += i; } } }
至于为什么,或许是因为编译器的优化。这里我们要专注于我们一开始的问题:Preemption Context Switches从何而来。从执行时间而来。当然这只是针对多线程间无锁情况,下面给它加上锁,看看是否有哪个因素也会影响到Preemption Context Switches。
实验二:多线程加锁
speedup-example-mutex-only.cpp
#include <pthread.h>
#include <stdio.h>
#include <stdlib.h>
#include <errno.h>
#include <assert.h>
#define N 4
#define M 30000
int nwait = 0;
volatile long long sum;
long loops = 6e3;
pthread_mutex_t mutex;
void set_affinity(int core_id) {
cpu_set_t cpuset;
CPU_ZERO(&cpuset);
CPU_SET(core_id, &cpuset);
assert(pthread_setaffinity_np(pthread_self(), sizeof(cpu_set_t), &cpuset) == 0);
}
void* thread_func(void *arg) {
set_affinity((int)(long)arg);
for (int j = 0; j < M; j++) {
phtread_mutex_lock(&mutex);
nwait++;
for (long i = 0; i < loops; i++)
sum += i;
phtread_mutex_unlock(&mutex);
for (long i = 0; i < loops; i++)
sum += i*i*i*i*i*i;
}
}
int main(int argc, char *argv[]) {
set_affinity(23);
pthread_t th[N];
int ret;
for(unsigned i=0; i<N; ++i) {
ret = pthread_create(&th[i], NULL, thread_func, (void*)i);
assert(!ret && "pthread_create() failed!");
}
for(unsigned i=0; i<N; ++i)
pthread_join(th[i], NULL);
exit(0);
}
VTune现象:
接下来我们改变线程数,即N等于8:(我们期望Unknown处的Preemption增加类似线性,而clone处的增加幅度大,即与多线程无锁的情况类似)
Unkown stack frame(s)的对Preemption Context Switches的贡献率任然不如clone。且在同等数目线程下,加锁情况下的clone要比不加锁的制造更多的Preemption Context Switches。如果用我们上面的“时间理论”来解释——加锁的运行时间明显比不加锁要多,也能解释,不过这并不充分,让我们执行以下验证:
void* thread_func(void *arg) { set_affinity((int)(long)arg); for (int j = 0; j < M; j++) { phtread_mutex_lock(&mutex); nwait++; for (long i = 0; i < loops; i++) sum += i; phtread_mutex_unlock(&mutex); } }
void* thread_func(void *arg) { set_affinity((int)(long)arg); for (int j = 0; j < M; j++) { phtread_mutex_lock(&mutex); nwait++; phtread_mutex_unlock(&mutex); for (long i = 0; i < loops; i++) sum += i*i*i*i*i*i; } }
void* thread_func(void *arg) { set_affinity((int)(long)arg); for (int j = 0; j < M; j++) { phtread_mutex_lock(&mutex); nwait++; phtread_mutex_unlock(&mutex); } }
void* thread_func(void *arg) { set_affinity((int)(long)arg); for (int j = 0; j < M; j++) { phtread_mutex_lock(&mutex); nwait++; for (long i = 0; i < loops; i++) sum += i; for (long i = 0; i < loops; i++) sum += i; phtread_mutex_unlock(&mutex); } }
void* thread_func(void *arg) { set_affinity((int)(long)arg); for (int j = 0; j < M; j++) { phtread_mutex_lock(&mutex); nwait++; for (long i = 0; i < loops; i++) sum += i; for (long i = 0; i < loops; i++) sum += i; for (long i = 0; i < loops; i++) sum += i; for (long i = 0; i < loops; i++) sum += i; phtread_mutex_unlock(&mutex); } }
void* thread_func(void *arg) { set_affinity((int)(long)arg); for (int j = 0; j < M; j++) { phtread_mutex_lock(&mutex); nwait++; for (long i = 0; i < loops; i++) { sum += i; sum += i; sum += i; sum += i; } phtread_mutex_unlock(&mutex); } }
我们发现,基本上加锁情况与无锁情况一致,不过我们还需做以下验证:
void* thread_func(void *arg) { set_affinity((int)(long)arg); for (int j = 0; j < M; j++) { phtread_mutex_lock(&mutex); nwait++; phtread_mutex_unlock(&mutex); for (long i = 0; i < loops; i++) sum += i*i*i*i*i*i; for (long i = 0; i < loops; i++) sum += i*i*i*i*i*i; } }
void* thread_func(void *arg) { set_affinity((int)(long)arg); for (int j = 0; j < M; j++) { phtread_mutex_lock(&mutex); nwait++; phtread_mutex_unlock(&mutex); for (long i = 0; i < loops; i++) sum += i*i*i*i*i*i; for (long i = 0; i < loops; i++) sum += i*i*i*i*i*i; for (long i = 0; i < loops; i++) sum += i*i*i*i*i*i; for (long i = 0; i < loops; i++) sum += i*i*i*i*i*i; } }
void* thread_func(void *arg) { set_affinity((int)(long)arg); for (int j = 0; j < M; j++) { phtread_mutex_lock(&mutex); nwait++; phtread_mutex_unlock(&mutex); for (long i = 0; i < loops; i++) { sum += i*i*i*i*i*i; sum += i*i*i*i*i*i; sum += i*i*i*i*i*i; sum += i*i*i*i*i*i; } } }
果然,在一定误差可容忍下,for循环是不区别加锁for和不加锁for,它们取得的效果基本一样——随着第二层for的数目增加,clone处的Preemption在增加;不过这里,单个for中增加子句的效果和增加for数目的效果基本一样,这与无锁是不同的;而且,有一个比较重要的差别:
在无锁的情况下,
A
void* thread_func(void *arg) {
set_affinity((int)(long)arg);
for (int j = 0; j < M; j++) {
nwait++;
for (long i = 0; i < loops; i++)
sum += i;
for (long i = 0; i < loops; i++)
sum += i*i*i*i*i*i;
}
}
和
B
void* thread_func(void *arg) {
set_affinity((int)(long)arg);
for (int j = 0; j < M; j++) {
nwait++;
for (long i = 0; i < loops; i++)
sum += i;
for (long i = 0; i < loops; i++)
sum += i;
}
}
clone处Preemption的数目基本一致,但在加锁的情况下:
C
void* thread_func(void *arg) {
set_affinity((int)(long)arg);
for (int j = 0; j < M; j++) {
phtread_mutex_lock(&mutex);
nwait++;
for (long i = 0; i < loops; i++)
sum += i;
phtread_mutex_unlock(&mutex);
for (long i = 0; i < loops; i++)
sum += i*i*i*i*i*i;
}
}
和
D
void* thread_func(void *arg) {
set_affinity((int)(long)arg);
for (int j = 0; j < M; j++) {
phtread_mutex_lock(&mutex);
nwait++;
for (long i = 0; i < loops; i++)
sum += i;
for (long i = 0; i < loops; i++)
sum += i;
phtread_mutex_unlock(&mutex);
}
}
clone处Preemption的数目不一样,前者要明显多于后者。但是如果我们将后者改为:
E
void* thread_func(void *arg) {
set_affinity((int)(long)arg);
for (int j = 0; j < M; j++) {
phtread_mutex_lock(&mutex);
nwait++;
for (long i = 0; i < loops; i++)
sum += i;
phtread_mutex_unlock(&mutex);
for (long i = 0; i < loops; i++)
sum += i;
}
}
则VTune分析有:
这就和
C
效果基本一样了。而解释C、D、E三者之间的差异,或许也可以用我们的“时间理论”。执行三者:
C
D
E
虽然D的运行时比C和E稍小,但我们不能直接将无锁情况下的时间理论应用到加锁情况。在说明原因之前,先看另一个程序:
F
void* thread_func(void *arg) {
set_affinity((int)(long)arg);
for (int j = 0; j < M; j++) {
phtread_mutex_lock(&mutex);
nwait++;
phtread_mutex_unlock(&mutex);
for (long i = 0; i < loops; i++)
sum += i*i*i*i*i*i;
for (long i = 0; i < loops; i++)
sum += i*i*i*i*i*i;
}
}
和D在clone处拥有基本一样的Preemption数,但二者的运行时间却大不一样。
F
所以“运行时间不一样,clone处的Preemption数不一样”,在这里就不适用了。看来无锁和加锁还是有个重要差别的。我们都知道在无锁情况下,所有子线程并行执行,VTune中有如下调度:
我们通过大量的观察发现,对于每个线程,每相隔1s就会有一次Preemption Context Switches,所以无锁情况下,随着运行时间的增加,clone处的Preemption数会增多。其实“时间理论”也适用于加锁情况,那为什么会出现上面C、D、E的情况,以及D和F的情况?我们也从调度图入手:
C
D
F
其实加锁和无锁的“时间理论”的区别在于:加锁情况中的C和D(基本串行化),并不是每一个线程中每隔1s就有一个Preemption;而加锁情况中的F(拥有并行化),每个线程中每隔1s会有一个Preemption。
这样对于C和D,由于C的运行时较D长,其中包含的Preemption比D多;而F虽然运行时比D短,但每个线程中的Preemption汇总就会和D一样多。
最终我们得出结论:
Preemption Context Switches的来源是——
对于拥有并行化的程序,运行时间越长,Preemption Context Switches越多;对于加锁导致串行化的程序,运行时间越长,Preemption Context Switches越多;对于加锁仍保留并行化的程序,运行时间越长,Preemption Context Switches越多。