How does mmap improve file reading speed?

Question

Assuming the address space can cover the file, it appears to me that mmap simply allocates a chunk of memory as large as the file about to be read, and creates a 1-to-1 rela

Answer 1

I was curious about this so I tried benchmarking whole-file reads for files of sizes 1, 2, 4, 8, etc., once with mmap (M) and once with read (R) (theoretically one call with the fstat-ed size, but it would retry if that call returned a partial result). After the reading/mmaping, a byte of each mmaped/read page was accessed in an non-optimizable fashion.

Here's my results:

Size   M(µs)   R(µs)
1      9.5     4.2
2      10.8    4.5
4      8.4     3.8
8      8.6     3.8
16     7.3     4
32     7.8     3.5
64     8.3     3.9
128    9.2     4.6
256    8.6     4.7
512    10.6    5.1
1.0Ki  9.8     4.7
2.0Ki  10.1    5.4
4.0Ki  10.5    5.6
8.0Ki  10.4    6.9
16Ki   9.9     10
32Ki   14.4    12.8
64Ki   16.1    23.7
128Ki  28.1    41.1
256Ki  34.5    82.4
512Ki  57.9    154.6
1.0Mi  103.5   325.8
2.0Mi  188.5   919.8
4.0Mi  396.3   1963.2
8.0Mi  798.8   3885
16Mi   1611.4  7660.2
32Mi   3207.4  23040.2
64Mi   6712.1  84491.9

It appears read is about twice as fast up to about 16Ki. From then on, mmap starts winning big time (for 64MiB files by a factor of 12).

(Tested on Linux with 3.19 on my laptop, 10^4 repeated reads to the same file.)