关于缓存性能的一些问题计算机架构

如何解决关于缓存性能的一些问题计算机架构

有关 X5650 处理器的详细信息，请访问 https://www.cpu-world.com/CPUs/Xeon/Intel-Xeon%20X5650%20-%20AT80614004320AD%20(BX80614X5650).html

重要说明： L3缓存大小：12288KB 缓存线大小：64

考虑以下两个函数，它们都将数组中的值增加 100。

void incrementVector1(INT4* v,int n) {
                       for (int k = 0; k < 100; ++k) {
                           for (int i = 0; i < n; ++i) {
                               v[i] = v[i] + 1;
} }
}
                   void incrementVector2(INT4* v,int n) {
                       for (int i = 0; i < n; ++i) {
                           for (int k = 0; k < 100; ++k) {
                               v[i] = v[i] + 1;
} }
}

The following data collected using the perf utility captures runtime information for executing
each of these functions on the Intel Xeon X5650 processor for various data sizes. In this data: • the program vector1.bin executes the function incrementVector1;
• the program vector2.bin executes the function incrementVector2;
• the programs take a command line argument which sets the value of n;
• both programs begin by allocating an array of size n and initializing all elements to 0. • LLC-loads means “last level cache loads”,the number of accesses to L3;
• LLC-load-misses means “last level cache misses”,the number of L3 cache misses.
Runtime performance of vector1.bin.
Performance counter stats for ’./vector1.bin 1000000’:
230,070      LLC-loads
3,280        LLC-load-misses           #    1.43% of all LL-cache references
0.383542737 seconds time elapsed
Performance counter stats for ’./vector1.bin 3000000’:
669,884      LLC-loads
242,876      LLC-load-misses           #   36.26% of all LL-cache references
1.156663301 seconds time elapsed
Performance counter stats for ’./vector1.bin 5000000’:
1,234,031    LLC-loads
898,577      LLC-load-misses           #   72.82% of all LL-cache references
1.941832434 seconds time elapsed
Performance counter stats for ’./vector1.bin 7000000’:
1,620,026      LLC-loads
1,142,275      LLC-load-misses           #   70.51% of all LL-cache references
2.621428714 seconds time elapsed
Performance counter stats for ’./vector1.bin 9000000’:
2,068,028      LLC-loads
1,422,269      LLC-load-misses           #   68.77% of all LL-cache references
3.308037628 seconds time elapsed
8
Runtime performance of vector2.bin.
Performance counter stats for ’./vector2.bin 1000000’:
16,464     LLC-loads
1,168      LLC-load-misses            #   7.049% of all LL-cache references
0.319311959 seconds time elapsed
Performance counter stats for ’./vector2.bin 3000000’:
42,052      LLC-loads
17,027      LLC-load-misses           #   40.49% of all LL-cache references
0.954854798 seconds time elapsed
Performance counter stats for ’./vector2.bin 5000000’:
63,991      LLC-loads
38,459      LLC-load-misses           #   60.10% of all LL-cache references
1.593526338 seconds time elapsed
Performance counter stats for ’./vector2.bin 7000000’:
99,773      LLC-loads
56,481      LLC-load-misses           #   56.61% of all LL-cache references
2.198810471 seconds time elapsed
Performance counter stats for ’./vector2.bin 9000000’:
120,456     LLC-loads
76,951      LLC-load-misses           #   63.88% of all LL-cache references
2.772653964 seconds time elapsed

我有两个问题：

1. 考虑 vector1.bin 的缓存未命中率。在向量大小 1000000 和 5000000 之间，缓存未命中率急剧增加。缓存未命中率增加的原因是什么？
2. 考虑两个程序的缓存未命中率。请注意，对于任何特定的数组大小，两个程序之间的未命中率大致相等。这是为什么？