mmap() vs read() and write() Benchmark

Question Text

According to the times shown by benchmark_cp.sh, which of the two implementations is faster?

Question Answers

• the one using mmap()

• the one using read() and write()

• they are roughly equivalent

• cp is faster if it uses copy_file_range()

Feedback

In our case, running the script a few times results in the following running times:

student@os:/.../support/file-mappings$ ./benchmark_cp.sh
make: Nothing to be done for 'all'.
Benchmarking mmap_cp on a 1GB file...

real    0m30,597s
user    0m0,569s
sys     0m2,286s
Benchmarking cp on a 1 GB file...

real    0m36,012s
user    0m0,039s
sys     0m2,469s


student@os:/.../support/file-mappings$ ./benchmark_cp.sh
make: Nothing to be done for 'all'.
Benchmarking mmap_cp on a 1 GB file...

real    0m27,803s
user    0m0,590s
sys     0m2,114s
Benchmarking cp on a 1 GB file...

real    0m35,607s
user    0m0,033s
sys     0m2,564s

So it seems that using mmap() rather than read() and write() yields about a 15% increase in performance. However, you might get different results on your system. This depends on your storage device (SSD vs HDD) and its specific speed (like its RPM for an HDD). So the more conservative answer is to say that this depends on external aspects and that, in general, the 2 implementations are more or less equivalent.

If you want to know why there isn't much of a difference between the 2 implementations, check out this explanation.

However, some newer Linux systems use an updated cp that doesn't use read and write anymore, but instead uses zero-copy, by means of the copy_file_range() syscall. This implementation is likely going to be significantly faster than the one using mmap, as shown in the snippet below:

student@os:/.../support/file-mappings$ ./benchmark_cp.sh
make: Nothing to be done for 'all'.
Benchmarking mmap_cp on a 1 GB file...

real    0m1,443s
user    0m0,429s
sys     0m0,971s
Benchmarking cp on a 1 GB file...

real    0m0,821s
user    0m0,001s
sys     0m0,602s