This repository contains a microbenchmark that uses mmap(2) to load a file from the system. The microbenchmark follows the same methodologies as the syscall_timing microbenchmark tool in FreeBSD. It is used for testing NVDIMM's performance in Gen9 machines.

usage

$ cc -o mmap mmap.c
mmap [-i iterations] [-l loops] [-s seconds] -f filesize -p path

sample screenshot of microbenchmark

$ ./mmap -f 100 -p sample.txt
Clock resolution: 0.000001000
test	loop	time	iterations	per-iteration
mmap_read	0	1.005074000	102038056	0.000000009
mmap_read	1	1.002209000	97507225	0.000000010
mmap_read	2	1.005069000	101301978	0.000000009
mmap_read	3	1.001100000	98339145	0.000000010
mmap_read	4	1.001014000	98030109	0.000000010
mmap_read	5	1.000977000	98597120	0.000000010
mmap_read	6	1.001026000	100037525	0.000000010
mmap_read	7	1.002810000	99390593	0.000000010
mmap_read	8	1.004042000	101932768	0.000000009
mmap_read	9	1.003278000	101852432	0.000000009

results:

The microbenchmark was tried with a 32GiB file on normal filesystem, and a 1GB file on tmpfs and nvdimm. The mmap-run script runs the executable with a size-range from 2KiBs to 1GiB for tmpfs and nvdimm, and up to 32GiBs for the file on the normal file system.

Since the microbenchmark strides randomly through the mmaped region, for bigger sizes, there are more cache misses, and thus the average read time increases. This is not, however, measuring specific cache level's access times. The effect of higher-level caches is more apparent as the mapped region's size increases.