我在Intel机器上观察到一些非常奇怪的读写访问性能。
我编写了一个首先分配数组的C程序。该程序的代码是[1];您可以通过运行Make来编译它。 （我不使用任何编译优化。）
程序的操作顺序如下：
allocate a char array;
init each element of array to be 1;
use clflush to flush the whole array from cache;
read each cache line of the array by using tmp = array[i];
(Do simple calculation after reading each cache line)
use clflush to flush the whole array from cache;
write each cache line of the array by using array[i] = tmp;
(Do the same simple calculation after reading each cache line)
我在英特尔（R）Xeon（R）CPU E5-1650 v2 @ 3.50GHz（Haswell arch。）上运行程序，禁用了涡轮增压。
我用来运行程序的命令是： sudo ./rw-latency-test-compute 5210 10 1
我得到整个阵列的读取延迟为6670us，而整个阵列的写入延迟为3518us。
有趣的部分是 如果我在读/写缓存行后没有进行任何计算，整个数组的读取延迟为2175us，而整个数组的写入延迟为3687us。
所以做计算似乎加快了执行......： - （
你对这种奇怪的表现有什么建议/解释吗？
程序的整个汇编代码可以在[2]找到。
内循环的汇编代码如下：
0000000000400898 <read_array>:
400898: 55 push %rbp
400899: 48 89 e5 mov %rsp,%rbp
40089c: 53 push %rbx
40089d: 48 83 ec 28 sub $0x28,%rsp
4008a1: 48 89 7d d8 mov %rdi,-0x28(%rbp)
4008a5: 48 89 75 d0 mov %rsi,-0x30(%rbp)
4008a9: c7 45 e8 00 00 00 00 movl $0x0,-0x18(%rbp)
4008b0: c7 45 e4 00 00 00 00 movl $0x0,-0x1c(%rbp)
4008b7: eb 58 jmp 400911 <read_array+0x79>
4008b9: b8 00 00 00 00 mov $0x0,%eax
4008be: e8 38 ff ff ff callq 4007fb <sw_barrier>
4008c3: 8b 45 e4 mov -0x1c(%rbp),%eax
4008c6: 48 98 cltq
4008c8: 48 03 45 d8 add -0x28(%rbp),%rax
4008cc: 0f b6 00 movzbl (%rax),%eax
4008cf: 88 45 ef mov %al,-0x11(%rbp)
4008d2: 0f be 45 ef movsbl -0x11(%rbp),%eax
4008d6: 89 c1 mov %eax,%ecx
4008d8: 03 4d e8 add -0x18(%rbp),%ecx
4008db: ba 01 80 00 80 mov $0x80008001,%edx
4008e0: 89 c8 mov %ecx,%eax
4008e2: f7 ea imul %edx
4008e4: 8d 04 0a lea (%rdx,%rcx,1),%eax
4008e7: 89 c2 mov %eax,%edx
4008e9: c1 fa 0f sar $0xf,%edx
4008ec: 89 c8 mov %ecx,%eax
4008ee: c1 f8 1f sar $0x1f,%eax
4008f1: 89 d3 mov %edx,%ebx
4008f3: 29 c3 sub %eax,%ebx
4008f5: 89 d8 mov %ebx,%eax
4008f7: 89 45 e8 mov %eax,-0x18(%rbp)
4008fa: 8b 55 e8 mov -0x18(%rbp),%edx
4008fd: 89 d0 mov %edx,%eax
4008ff: c1 e0 10 shl $0x10,%eax
400902: 29 d0 sub %edx,%eax
400904: 89 ca mov %ecx,%edx
400906: 29 c2 sub %eax,%edx
400908: 89 d0 mov %edx,%eax
40090a: 89 45 e8 mov %eax,-0x18(%rbp)
40090d: 83 45 e4 40 addl $0x40,-0x1c(%rbp)
400911: 8b 45 e4 mov -0x1c(%rbp),%eax
400914: 48 98 cltq
400916: 48 3b 45 d0 cmp -0x30(%rbp),%rax
40091a: 7c 9d jl 4008b9 <read_array+0x21>
40091c: b8 e1 0f 40 00 mov $0x400fe1,%eax
400921: 8b 55 e8 mov -0x18(%rbp),%edx
400924: 89 d6 mov %edx,%esi
400926: 48 89 c7 mov %rax,%rdi
400929: b8 00 00 00 00 mov $0x0,%eax
40092e: e8 3d fd ff ff callq 400670 <printf@plt>
400933: 48 83 c4 28 add $0x28,%rsp
400937: 5b pop %rbx
400938: 5d pop %rbp
400939: c3 retq
000000000040093a <write_array>:
40093a: 55 push %rbp
40093b: 48 89 e5 mov %rsp,%rbp
40093e: 53 push %rbx
40093f: 48 83 ec 28 sub $0x28,%rsp
400943: 48 89 7d d8 mov %rdi,-0x28(%rbp)
400947: 48 89 75 d0 mov %rsi,-0x30(%rbp)
40094b: c6 45 ef 01 movb $0x1,-0x11(%rbp)
40094f: c7 45 e8 00 00 00 00 movl $0x0,-0x18(%rbp)
400956: c7 45 e4 00 00 00 00 movl $0x0,-0x1c(%rbp)
40095d: eb 63 jmp 4009c2 <write_array+0x88>
40095f: b8 00 00 00 00 mov $0x0,%eax
400964: e8 92 fe ff ff callq 4007fb <sw_barrier>
400969: 8b 45 e4 mov -0x1c(%rbp),%eax
40096c: 48 98 cltq
40096e: 48 03 45 d8 add -0x28(%rbp),%rax
400972: 0f b6 55 ef movzbl -0x11(%rbp),%edx
400976: 88 10 mov %dl,(%rax)
400978: 8b 45 e4 mov -0x1c(%rbp),%eax
40097b: 48 98 cltq
40097d: 48 03 45 d8 add -0x28(%rbp),%rax
400981: 0f b6 00 movzbl (%rax),%eax
400984: 0f be c0 movsbl %al,%eax
400987: 89 c1 mov %eax,%ecx
400989: 03 4d e8 add -0x18(%rbp),%ecx
40098c: ba 01 80 00 80 mov $0x80008001,%edx
400991: 89 c8 mov %ecx,%eax
400993: f7 ea imul %edx
400995: 8d 04 0a lea (%rdx,%rcx,1),%eax
400998: 89 c2 mov %eax,%edx
40099a: c1 fa 0f sar $0xf,%edx
40099d: 89 c8 mov %ecx,%eax
40099f: c1 f8 1f sar $0x1f,%eax
4009a2: 89 d3 mov %edx,%ebx
4009a4: 29 c3 sub %eax,%ebx
4009a6: 89 d8 mov %ebx,%eax
4009a8: 89 45 e8 mov %eax,-0x18(%rbp)
4009ab: 8b 55 e8 mov -0x18(%rbp),%edx
4009ae: 89 d0 mov %edx,%eax
4009b0: c1 e0 10 shl $0x10,%eax
4009b3: 29 d0 sub %edx,%eax
4009b5: 89 ca mov %ecx,%edx
4009b7: 29 c2 sub %eax,%edx
4009b9: 89 d0 mov %edx,%eax
4009bb: 89 45 e8 mov %eax,-0x18(%rbp)
4009be: 83 45 e4 40 addl $0x40,-0x1c(%rbp)
4009c2: 8b 45 e4 mov -0x1c(%rbp),%eax
4009c5: 48 98 cltq
4009c7: 48 3b 45 d0 cmp -0x30(%rbp),%rax
4009cb: 7c 92 jl 40095f <write_array+0x25>
4009cd: b8 ee 0f 40 00 mov $0x400fee,%eax
4009d2: 8b 55 e8 mov -0x18(%rbp),%edx
4009d5: 89 d6 mov %edx,%esi
4009d7: 48 89 c7 mov %rax,%rdi
4009da: b8 00 00 00 00 mov $0x0,%eax
4009df: e8 8c fc ff ff callq 400670 <printf@plt>
4009e4: 48 83 c4 28 add $0x28,%rsp
4009e8: 5b pop %rbx
4009e9: 5d pop %rbp
4009ea: c3 retq
[1] https://github.com/PennPanda/rw-latency-test/blob/master/rw-latency-test-compute.c
答案 0 :(得分：4)
写入比读取更快，因为如果您从RAM读取并使用该值（即，您不必读取和丢弃），则处理器必须在使用该值时停止读取。但是，写入异步进行并且永远不会停止。