两个不同内核上的相同计算密集型函数 运行 导致不同的延迟
Same compute intensive function running on two different cores resulting in different latency
#include <pthread.h>
#include <fcntl.h>
#include <unistd.h>
#include <iostream>
#include <chrono>
using namespace std;
static inline void stick_this_thread_to_core(int core_id);
static inline void* incrementLoop(void* arg);
struct BenchmarkData {
long long iteration_count;
int core_id;
};
pthread_barrier_t g_barrier;
int main(int argc, char** argv)
{
if(argc != 3) {
cout << "Usage: ./a.out <core_id> <core_id>" << endl;
return EXIT_FAILURE;
}
cout << "================================================ STARTING ================================================" << endl;
int core1 = std::stoi(argv[1]);
int core2 = std::stoi(argv[2]);
pthread_barrier_init(&g_barrier, nullptr, 2);
const long long iteration_count = 100'000'000'000;
BenchmarkData benchmark_data1{iteration_count, core1};
BenchmarkData benchmark_data2{iteration_count, core2};
pthread_t worker1, worker2;
pthread_create(&worker1, nullptr, incrementLoop, static_cast<void*>(&benchmark_data1));
cout << "Created worker1" << endl;
pthread_create(&worker2, nullptr, incrementLoop, static_cast<void*>(&benchmark_data2));
cout << "Created worker2" << endl;
pthread_join(worker1, nullptr);
cout << "Joined worker1" << endl;
pthread_join(worker2, nullptr);
cout << "Joined worker2" << endl;
return EXIT_SUCCESS;
}
static inline void stick_this_thread_to_core(int core_id) {
int num_cores = sysconf(_SC_NPROCESSORS_ONLN);
if (core_id < 0 || core_id >= num_cores) {
cerr << "Core " << core_id << " is out of assignable range.\n";
return;
}
cpu_set_t cpuset;
CPU_ZERO(&cpuset);
CPU_SET(core_id, &cpuset);
pthread_t current_thread = pthread_self();
int res = pthread_setaffinity_np(current_thread, sizeof(cpu_set_t), &cpuset);
if(res == 0) {
cout << "Thread bound to core " << core_id << " successfully." << endl;
} else {
cerr << "Error in binding this thread to core " << core_id << '\n';
}
}
static inline void* incrementLoop(void* arg)
{
BenchmarkData* arg_ = static_cast<BenchmarkData*>(arg);
int core_id = arg_->core_id;
long long iteration_count = arg_->iteration_count;
stick_this_thread_to_core(core_id);
cout << "Thread bound to core " << core_id << " will now wait for the barrier." << endl;
pthread_barrier_wait(&g_barrier);
cout << "Thread bound to core " << core_id << " is done waiting for the barrier." << endl;
long long data = 0;
long long i;
cout << "Thread bound to core " << core_id << " will now increment private data " << iteration_count / 1'000'000'000.0 << " billion times." << endl;
std::chrono::steady_clock::time_point begin = std::chrono::steady_clock::now();
for(i = 0; i < iteration_count; ++i) {
++data;
__asm__ volatile("": : :"memory");
}
std::chrono::steady_clock::time_point end = std::chrono::steady_clock::now();
unsigned long long elapsed_time = std::chrono::duration_cast<std::chrono::milliseconds>(end - begin).count();
cout << "Elapsed time: " << elapsed_time << " ms, core: " << core_id << ", iteration_count: " << iteration_count << ", data value: " << data << ", i: " << i << endl;
return nullptr;
}
我的编译命令：
g++ main.cpp -Ofast -lpthread
我有一个基本的微基准测试代码，其中我从主线程生成了两个工作线程。我将生成的线程绑定到核心 0 和 1。两个线程都调用一个名为 incrementLoop 的函数。在此函数中，线程将在 for 循环中将局部（每个线程私有）递增 1000 亿次。然后我测量两个线程在 for 循环中花费的时间。但是，两个不同内核上的线程 运行ning 在我的 Skylake 服务器上产生截然不同的经过时间结果，但在我的 Cascade Lake 服务器上经过时间非常接近。这可能是什么原因？
这是 Cascade Lake 的输出，但是由于我没有保护输出流免受相互访问，由于竞争，输出有点歪：
Created worker1
Created worker2
Thread bound to core Thread bound to core 0 successfully.
Thread bound to core 10 will now wait for the barrier. successfully.
Thread bound to core 1 will now wait for the barrier.
Thread bound to core 1 is done waiting for the barrier.
Thread bound to core 1 will now increment private data 100 billion times.
Thread bound to core 0 is done waiting for the barrier.
Thread bound to core 0 will now increment private data 100 billion times.
Elapsed time: 20861 ms, core: 1, iteration_count: 100000000000, data value: 100000000000, i: 100000000000
Elapsed time: 20925 ms, core: 0, iteration_count: 100000000000, data value: 100000000000, i: 100000000000
Joined worker1
Joined worker2
这是 Skylake 的输出，这让我很困惑：
================================================ STARTING ================================================
Created worker1
Created worker2
Thread bound to core Thread bound to core 01 successfully. successfully.
Thread bound to core 0 will now wait for the barrier.
Thread bound to core 1 will now wait for the barrier.
Thread bound to core 1 is done waiting for the barrier.
Thread bound to core 1 will now increment private data Thread bound to core 0 is done waiting for the barrier.
Thread bound to core 0 will now increment private data 100100 billion times. billion times.
Elapsed time: 27243 ms, core: 0, iteration_count: 100000000000, data value: 100000000000, i: 100000000000
Joined worker1
Elapsed time: 83536 ms, core: 1, iteration_count: 100000000000, data value: 100000000000, i: 100000000000
Joined worker2
Skylake 规格：
Architecture: x86_64
CPU op-mode(s): 32-bit, 64-bit
Byte Order: Little Endian
CPU(s): 18
On-line CPU(s) list: 0-17
Thread(s) per core: 1
Core(s) per socket: 18
Socket(s): 1
NUMA node(s): 1
Vendor ID: GenuineIntel
CPU family: 6
Model: 85
Model name: Intel(R) Xeon(R) Gold 6154 CPU @ 3.00GHz
Stepping: 4
CPU MHz: 1796.630
CPU max MHz: 3700.0000
CPU min MHz: 1200.0000
BogoMIPS: 6000.00
Virtualization: VT-x
L1d cache: 32K
L1i cache: 32K
L2 cache: 1024K
L3 cache: 25344K
NUMA node0 CPU(s): 0-17
Flags: fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush dts acpi mmx fxsr sse sse2 ss ht tm pbe syscall nx pdpe1gb rdtscp lm constant_tsc art arch_perfmon pebs bts rep_good nopl xtopology nonstop_tsc aperfmperf eagerfpu pni pclmulqdq dtes64 monitor ds_cpl vmx smx est tm2 ssse3 sdbg fma cx16 xtpr pdcm pcid dca sse4_1 sse4_2 x2apic movbe popcnt tsc_deadline_timer aes xsave avx f16c rdrand lahf_lm abm 3dnowprefetch epb cat_l3 cdp_l3 invpcid_single intel_ppin intel_pt ssbd mba ibrs ibpb stibp tpr_shadow vnmi flexpriority ept vpid fsgsbase tsc_adjust bmi1 hle avx2 smep bmi2 erms invpcid rtm cqm mpx rdt_a avx512f avx512dq rdseed adx smap clflushopt clwb avx512cd avx512bw avx512vl xsaveopt xsavec xgetbv1 cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local dtherm ida arat pln pts pku ospke md_clear spec_ctrl intel_stibp flush_l1d
喀斯喀特湖规格：
Architecture: x86_64
CPU op-mode(s): 32-bit, 64-bit
Byte Order: Little Endian
CPU(s): 18
On-line CPU(s) list: 0-17
Thread(s) per core: 1
Core(s) per socket: 18
Socket(s): 1
NUMA node(s): 1
Vendor ID: GenuineIntel
CPU family: 6
Model: 85
Model name: Intel(R) Core(TM) i9-10980XE CPU @ 3.00GHz
Stepping: 7
CPU MHz: 4799.926
CPU max MHz: 4800.0000
CPU min MHz: 1200.0000
BogoMIPS: 6000.00
Virtualization: VT-x
L1d cache: 32K
L1i cache: 32K
L2 cache: 1024K
L3 cache: 25344K
NUMA node0 CPU(s): 0-17
Flags: fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush dts acpi mmx fxsr sse sse2 ss ht tm pbe syscall nx pdpe1gb rdtscp lm constant_tsc art arch_perfmon pebs bts rep_good nopl xtopology nonstop_tsc aperfmperf eagerfpu pni pclmulqdq dtes64 monitor ds_cpl vmx est tm2 ssse3 sdbg fma cx16 xtpr pdcm pcid dca sse4_1 sse4_2 x2apic movbe popcnt tsc_deadline_timer aes xsave avx f16c rdrand lahf_lm abm 3dnowprefetch epb cat_l3 cdp_l3 invpcid_single intel_pt ssbd mba ibrs ibpb stibp ibrs_enhanced tpr_shadow vnmi flexpriority ept vpid fsgsbase tsc_adjust bmi1 hle avx2 smep bmi2 erms invpcid rtm cqm mpx rdt_a avx512f avx512dq rdseed adx smap clflushopt clwb avx512cd avx512bw avx512vl xsaveopt xsavec xgetbv1 cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local dtherm ida arat pln pts avx512_vnni md_clear spec_ctrl intel_stibp flush_l1d arch_capabilities
另请注意，我在两台服务器上都隔离了核心：
cat /proc/cmdline 对于 Skylake：
BOOT_IMAGE=/vmlinuz-3.10.0-1160.21.1.el7.x86_64 root=/dev/mapper/centos-root ro crashkernel=auto rd.lvm.lv=centos/root rd.lvm.lv=centos/swap rhgb processor.max_cstate=0 nohz_full=0-15 iommu=off intel_idle.max_cstate=0 isolcpus=0-15 idle=poll nosoftlockup mce=ignore_ce quiet skew_tick=1
cat /proc/cmdline 喀斯喀特湖：
BOOT_IMAGE=/vmlinuz-3.10.0-1160.45.1.el7.x86_64 root=/dev/mapper/centos-root ro crashkernel=auto spectre_v2=retpoline rd.lvm.lv=centos/root rd.lvm.lv=centos/swap rhgb quiet nosoftlockup mce=ignore_ce intel_idle.max_cstate=0 pcie_aspm=performance ipmi_si.force_kipmi=0 nmi_watchdog=0 hpet=disabled noht processor.max_cstate=0 nohalt isolcpus=2-17 nohz=on nohz_full=2-17 rcu_nocbs=2-17 iommu=off audit=0 idle=poll skew_tick=1 skew_tick=1 skew_tick=1 skew_tick=1
我多次 运行 这些基准测试，每个 运行 的结果都相同或彼此接近。两台服务器上都没有其他进程 运行ning。
事实证明，核心 0、16、17 在我的 Skylake 服务器上 运行 的频率要高得多。
#include <pthread.h>
#include <fcntl.h>
#include <unistd.h>
#include <iostream>
#include <chrono>
using namespace std;
static inline void stick_this_thread_to_core(int core_id);
static inline void* incrementLoop(void* arg);
struct BenchmarkData {
long long iteration_count;
int core_id;
};
pthread_barrier_t g_barrier;
int main(int argc, char** argv)
{
if(argc != 3) {
cout << "Usage: ./a.out <core_id> <core_id>" << endl;
return EXIT_FAILURE;
}
cout << "================================================ STARTING ================================================" << endl;
int core1 = std::stoi(argv[1]);
int core2 = std::stoi(argv[2]);
pthread_barrier_init(&g_barrier, nullptr, 2);
const long long iteration_count = 100'000'000'000;
BenchmarkData benchmark_data1{iteration_count, core1};
BenchmarkData benchmark_data2{iteration_count, core2};
pthread_t worker1, worker2;
pthread_create(&worker1, nullptr, incrementLoop, static_cast<void*>(&benchmark_data1));
cout << "Created worker1" << endl;
pthread_create(&worker2, nullptr, incrementLoop, static_cast<void*>(&benchmark_data2));
cout << "Created worker2" << endl;
pthread_join(worker1, nullptr);
cout << "Joined worker1" << endl;
pthread_join(worker2, nullptr);
cout << "Joined worker2" << endl;
return EXIT_SUCCESS;
}
static inline void stick_this_thread_to_core(int core_id) {
int num_cores = sysconf(_SC_NPROCESSORS_ONLN);
if (core_id < 0 || core_id >= num_cores) {
cerr << "Core " << core_id << " is out of assignable range.\n";
return;
}
cpu_set_t cpuset;
CPU_ZERO(&cpuset);
CPU_SET(core_id, &cpuset);
pthread_t current_thread = pthread_self();
int res = pthread_setaffinity_np(current_thread, sizeof(cpu_set_t), &cpuset);
if(res == 0) {
cout << "Thread bound to core " << core_id << " successfully." << endl;
} else {
cerr << "Error in binding this thread to core " << core_id << '\n';
}
}
static inline void* incrementLoop(void* arg)
{
BenchmarkData* arg_ = static_cast<BenchmarkData*>(arg);
int core_id = arg_->core_id;
long long iteration_count = arg_->iteration_count;
stick_this_thread_to_core(core_id);
cout << "Thread bound to core " << core_id << " will now wait for the barrier." << endl;
pthread_barrier_wait(&g_barrier);
cout << "Thread bound to core " << core_id << " is done waiting for the barrier." << endl;
long long data = 0;
long long i;
cout << "Thread bound to core " << core_id << " will now increment private data " << iteration_count / 1'000'000'000.0 << " billion times." << endl;
std::chrono::steady_clock::time_point begin = std::chrono::steady_clock::now();
for(i = 0; i < iteration_count; ++i) {
++data;
__asm__ volatile("": : :"memory");
}
std::chrono::steady_clock::time_point end = std::chrono::steady_clock::now();
unsigned long long elapsed_time = std::chrono::duration_cast<std::chrono::milliseconds>(end - begin).count();
cout << "Elapsed time: " << elapsed_time << " ms, core: " << core_id << ", iteration_count: " << iteration_count << ", data value: " << data << ", i: " << i << endl;
return nullptr;
}
我的编译命令：
g++ main.cpp -Ofast -lpthread
我有一个基本的微基准测试代码，其中我从主线程生成了两个工作线程。我将生成的线程绑定到核心 0 和 1。两个线程都调用一个名为 incrementLoop 的函数。在此函数中，线程将在 for 循环中将局部（每个线程私有）递增 1000 亿次。然后我测量两个线程在 for 循环中花费的时间。但是，两个不同内核上的线程 运行ning 在我的 Skylake 服务器上产生截然不同的经过时间结果，但在我的 Cascade Lake 服务器上经过时间非常接近。这可能是什么原因？
这是 Cascade Lake 的输出，但是由于我没有保护输出流免受相互访问，由于竞争，输出有点歪：
Created worker1
Created worker2
Thread bound to core Thread bound to core 0 successfully.
Thread bound to core 10 will now wait for the barrier. successfully.
Thread bound to core 1 will now wait for the barrier.
Thread bound to core 1 is done waiting for the barrier.
Thread bound to core 1 will now increment private data 100 billion times.
Thread bound to core 0 is done waiting for the barrier.
Thread bound to core 0 will now increment private data 100 billion times.
Elapsed time: 20861 ms, core: 1, iteration_count: 100000000000, data value: 100000000000, i: 100000000000
Elapsed time: 20925 ms, core: 0, iteration_count: 100000000000, data value: 100000000000, i: 100000000000
Joined worker1
Joined worker2
这是 Skylake 的输出，这让我很困惑：
================================================ STARTING ================================================
Created worker1
Created worker2
Thread bound to core Thread bound to core 01 successfully. successfully.
Thread bound to core 0 will now wait for the barrier.
Thread bound to core 1 will now wait for the barrier.
Thread bound to core 1 is done waiting for the barrier.
Thread bound to core 1 will now increment private data Thread bound to core 0 is done waiting for the barrier.
Thread bound to core 0 will now increment private data 100100 billion times. billion times.
Elapsed time: 27243 ms, core: 0, iteration_count: 100000000000, data value: 100000000000, i: 100000000000
Joined worker1
Elapsed time: 83536 ms, core: 1, iteration_count: 100000000000, data value: 100000000000, i: 100000000000
Joined worker2
Skylake 规格：
Architecture: x86_64
CPU op-mode(s): 32-bit, 64-bit
Byte Order: Little Endian
CPU(s): 18
On-line CPU(s) list: 0-17
Thread(s) per core: 1
Core(s) per socket: 18
Socket(s): 1
NUMA node(s): 1
Vendor ID: GenuineIntel
CPU family: 6
Model: 85
Model name: Intel(R) Xeon(R) Gold 6154 CPU @ 3.00GHz
Stepping: 4
CPU MHz: 1796.630
CPU max MHz: 3700.0000
CPU min MHz: 1200.0000
BogoMIPS: 6000.00
Virtualization: VT-x
L1d cache: 32K
L1i cache: 32K
L2 cache: 1024K
L3 cache: 25344K
NUMA node0 CPU(s): 0-17
Flags: fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush dts acpi mmx fxsr sse sse2 ss ht tm pbe syscall nx pdpe1gb rdtscp lm constant_tsc art arch_perfmon pebs bts rep_good nopl xtopology nonstop_tsc aperfmperf eagerfpu pni pclmulqdq dtes64 monitor ds_cpl vmx smx est tm2 ssse3 sdbg fma cx16 xtpr pdcm pcid dca sse4_1 sse4_2 x2apic movbe popcnt tsc_deadline_timer aes xsave avx f16c rdrand lahf_lm abm 3dnowprefetch epb cat_l3 cdp_l3 invpcid_single intel_ppin intel_pt ssbd mba ibrs ibpb stibp tpr_shadow vnmi flexpriority ept vpid fsgsbase tsc_adjust bmi1 hle avx2 smep bmi2 erms invpcid rtm cqm mpx rdt_a avx512f avx512dq rdseed adx smap clflushopt clwb avx512cd avx512bw avx512vl xsaveopt xsavec xgetbv1 cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local dtherm ida arat pln pts pku ospke md_clear spec_ctrl intel_stibp flush_l1d
喀斯喀特湖规格：
Architecture: x86_64
CPU op-mode(s): 32-bit, 64-bit
Byte Order: Little Endian
CPU(s): 18
On-line CPU(s) list: 0-17
Thread(s) per core: 1
Core(s) per socket: 18
Socket(s): 1
NUMA node(s): 1
Vendor ID: GenuineIntel
CPU family: 6
Model: 85
Model name: Intel(R) Core(TM) i9-10980XE CPU @ 3.00GHz
Stepping: 7
CPU MHz: 4799.926
CPU max MHz: 4800.0000
CPU min MHz: 1200.0000
BogoMIPS: 6000.00
Virtualization: VT-x
L1d cache: 32K
L1i cache: 32K
L2 cache: 1024K
L3 cache: 25344K
NUMA node0 CPU(s): 0-17
Flags: fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush dts acpi mmx fxsr sse sse2 ss ht tm pbe syscall nx pdpe1gb rdtscp lm constant_tsc art arch_perfmon pebs bts rep_good nopl xtopology nonstop_tsc aperfmperf eagerfpu pni pclmulqdq dtes64 monitor ds_cpl vmx est tm2 ssse3 sdbg fma cx16 xtpr pdcm pcid dca sse4_1 sse4_2 x2apic movbe popcnt tsc_deadline_timer aes xsave avx f16c rdrand lahf_lm abm 3dnowprefetch epb cat_l3 cdp_l3 invpcid_single intel_pt ssbd mba ibrs ibpb stibp ibrs_enhanced tpr_shadow vnmi flexpriority ept vpid fsgsbase tsc_adjust bmi1 hle avx2 smep bmi2 erms invpcid rtm cqm mpx rdt_a avx512f avx512dq rdseed adx smap clflushopt clwb avx512cd avx512bw avx512vl xsaveopt xsavec xgetbv1 cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local dtherm ida arat pln pts avx512_vnni md_clear spec_ctrl intel_stibp flush_l1d arch_capabilities
另请注意，我在两台服务器上都隔离了核心：
cat /proc/cmdline 对于 Skylake：
BOOT_IMAGE=/vmlinuz-3.10.0-1160.21.1.el7.x86_64 root=/dev/mapper/centos-root ro crashkernel=auto rd.lvm.lv=centos/root rd.lvm.lv=centos/swap rhgb processor.max_cstate=0 nohz_full=0-15 iommu=off intel_idle.max_cstate=0 isolcpus=0-15 idle=poll nosoftlockup mce=ignore_ce quiet skew_tick=1
cat /proc/cmdline 喀斯喀特湖：
BOOT_IMAGE=/vmlinuz-3.10.0-1160.45.1.el7.x86_64 root=/dev/mapper/centos-root ro crashkernel=auto spectre_v2=retpoline rd.lvm.lv=centos/root rd.lvm.lv=centos/swap rhgb quiet nosoftlockup mce=ignore_ce intel_idle.max_cstate=0 pcie_aspm=performance ipmi_si.force_kipmi=0 nmi_watchdog=0 hpet=disabled noht processor.max_cstate=0 nohalt isolcpus=2-17 nohz=on nohz_full=2-17 rcu_nocbs=2-17 iommu=off audit=0 idle=poll skew_tick=1 skew_tick=1 skew_tick=1 skew_tick=1
我多次 运行 这些基准测试，每个 运行 的结果都相同或彼此接近。两台服务器上都没有其他进程 运行ning。
事实证明，核心 0、16、17 在我的 Skylake 服务器上 运行 的频率要高得多。