Written here are performance issues found.
Timing stats of live migration of a small container with 11 tasks is
- Total time ~3.5 seconds
- Frozen time ~3.0 seconds
- Pre-dump stages ~0.5 seconds each
- Restore time ~1.9 seconds
- Images transfer time ~0.3 seconds
Below is the list of issues found
Images IO
Too many write-s and read-s. Each small object issues a write/read request into file. Need to improve.
Dump
Surprisingly, but the mem-drain time is not the biggest. It's "only" ~0.02 seconds. There are places in code that take longer.
parse_smaps
Time spent in this routine is up to 0.2 seconds on dump. This one exploits /proc heavily. For a container with 11 tasks the syscall stats look like
834 read 1451 fstat 1462 close 1642 openat
while opens and stats happen on
193 openat(4, "map-symlink", O_RDONLY) = -1 ENOENT (No such file or directory)
1438 openat(4, "map-symlink", O_RDONLY) = 5
11 openat(AT_FDCWD, "/proc/$pid/map_files", O_RDONLY|O_NONBLOCK|O_DIRECTORY|O_CLOEXEC) = 4
parasite_fuxup_vdso
The 2nd longest routine. It takes ~ 0.07 seconds on dump.
Too many readlink-s
Some files are readlinked 3-4 times in a row.
Restore
Fork vs VMA restore
We restore task's mappings before it goes forking to handle COW. This effectively serializes forking.
Restoring VMAs
There are 4 stages in VMA restore. Relative times of each are below
- Reading images 1%
- Mapping huuge premap area << 1%
- (Re-)mapping sub-areas 73%
- Filling area with data 26%
The 3rd stage has two parts. With timings:
- Opening filemap fd 85%
- Maping vma 15%
Parse maps instead of smaps
When parsing self maps for restorer hint, it's enough to read smaller /proc/self/maps