我正在研究缓存对 x86-64 CPU 性能的影响。我一直在使用 Linux 的 perf 来监控缓存命中/未命中率，特别是这些计数器：
mem_inst_retired.all_loadsmem_load_retired.l1_hitmem_load_retired.l1_miss我希望如此，因为加载指令可能会命中或错过 L1 缓存 - 没有其他选择，因为所有常规加载都会通过缓存。我已经阅读了英特尔文档
here，据我所知，其中并没有说太多内容，也没有任何内容可以反驳我的想法。 但是，当运行某些代码时，我注意到总和远低于
all_loads ~= l1_hit + l1_miss。
例如，以下程序集将 10 亿个内存位置与步长 32B 相加：
all_loads
它产生以下性能结果：
.intel_syntax noprefix
.globl main
main:
sub rsp, 8 # Stack alignment for call
push r12
push r13
mov r12, 0xfffffff # Array size
mov rdi, r12
call malloc # Allocate array
mov r13, rax # Save array pointer
mov rdi, r13 # Write to array to force page-in
mov rsi, 42
mov rdx, r12
call memset
mov rcx, 1000000000 # Loop counter
mov rdx, 0 # Index sequence start
mov rax, 0 # Result accumulator
.p2align 4 # Skylake JCC alignment issue
loop:
mov rdi, rdx
and rdi, r12 # Mask index to array size
movzx rsi, BYTE PTR [r13+rdi] # Read from array index
add rax, rsi
lea rdx, [rdx+32] # Generate next array index
dec rcx # Loop counter & condiiton
jnz loop
pop r13
pop r12
add rsp, 8
ret
~$ perf stat -e instructions,cycles,mem_inst_retired.all_loads,mem_load_retired.l1_hit,mem_load_retired.l1_miss ./test
Performance counter stats for './test':
7,000,215,742 instructions:u # 1.25 insn per cycle
5,589,048,737 cycles:u
998,622,922 mem_inst_retired.all_loads:u
17,215,080 mem_load_retired.l1_hit:u
424,118,595 mem_load_retired.l1_miss:u
1.939187022 seconds time elapsed
1.826889000 seconds user
0.112177000 seconds sys
大约为 10 亿（尽管有点低，但可能是一些采样人工制品）。然而
all_loadsl1_hit + l1_miss 和
l1_hitl1_missall_loads。只有在中间地带，平等才会被打破。
编辑：我在两个 CPU 上进行了测试：Kaby Lake 和 Ice Lake。两者显示相同的结果。
，而不是分配新的缓存行。事实证明，这既不是 all_loads ~= l1_hit + l1_miss 也不是
l1_hitl1_missmem_load_retired.fb_hit在我的 Skylake i7-6700k 上使用此测试程序，
l1_miss仅比
mem_inst_retired.all_loadsmem_load_retired.fb_hit + mem_load_retired.l1_hit + mem_load_retired.l1_miss计数的内容，而不是三个更具体的计数器中的任何一个，仍然存在一些谜团。我原以为它们会更接近完全相等，特别是对于
mem_inst_retired.all_loads--all-user:u可以轻松地将数字复制/粘贴到
perf stat --no-big-num --all-user -e ...calc
我的 EPP
$ perf stat --all-user --no-big-num -e task-clock,page-faults,instructions,cycles,mem_inst_retired.all_loads,mem_load_retired.l1_hit,mem_load_retired.l1_miss,mem_load_retired.fb_hit ./a.out
Performance counter stats for './a.out':
1673.21 msec task-clock # 0.997 CPUs utilized
183 page-faults # 109.371 /sec
7000141672 instructions # 1.56 insn per cycle
4475942186 cycles # 2.675 GHz
999892622 mem_inst_retired.all_loads # 597.590 M/sec
10563966 mem_load_retired.l1_hit # 6.314 M/sec
449822478 mem_load_retired.l1_miss # 268.838 M/sec
533816318 mem_load_retired.fb_hit # 319.038 M/sec
1.677680356 seconds time elapsed
1.647640000 seconds user
0.023197000 seconds sys
$ perf stat --all-user --no-big-num -e task-clock,page-faults,instructions,cycles,mem_inst_retired.all_loads,mem_load_retired.l1_hit,mem_load_retired.l1_miss,mem_load_retired.fb_hit ./a.out
Performance counter stats for './a.out':
1649.17 msec task-clock # 1.000 CPUs utilized
182 page-faults # 110.359 /sec
7000141486 instructions # 1.58 insn per cycle
4419739785 cycles # 2.680 GHz
999850616 mem_inst_retired.all_loads # 606.275 M/sec
9372903 mem_load_retired.l1_hit # 5.683 M/sec
450146459 mem_load_retired.l1_miss # 272.953 M/sec
534244404 mem_load_retired.fb_hit # 323.947 M/sec
1.649504255 seconds time elapsed
1.634275000 seconds user
0.013270000 seconds sys
）设置为
/sys/devices/system/cpu/cpufreq/policy*/energy_performance_preferencebalance-performance