On 8.6.2014, at 11.59, Ralf Hildebrandt r@sys4.de wrote:
I observed several long running dovecot/lmtp processes hogging the CPU. I then strace'd them (strace -c -p 6375) and found them munmap()-ing a lot:
% time seconds usecs/call calls errors syscall
97.18 19.592537 1878 10430 munmap 2.28 0.458984 36 12696 epoll_ctl 0.26 0.052926 10 5288 fdatasync 0.21 0.042472 3 13679 epoll_wait ... snip ...
Why would that happen? (dovecot 2.1.17)
Difficult to say. It could be munmap()ing memory allocations or it could be munmap()ing Dovecot index files. Weren't there equivalent number of mmap() calls?..
BTW. In v2.2 the index file handling is faster for LDA/LMTP because it doesn't even try to mmap() the full indexes into memory.