John Morrissey wrote:
2.4.16 (with BDB4.7 + patches and the POSIX threads build options) has been stable on amd64/x86_64 for us, but I'm encountering problems with 2.4.16 on i386.
On what OS?
I've been seeing similar issues on Linux, but using heap profilers / leak checkers don't turn up any results. At least in the Linux case, it appears to be general heap fragmentation caused by continual mmap/munmap of malloc'd memory blocks.
slapd will run happily for several hours at a RSS of ~1.2GB, but then will suddenly gobble memory until it hits the i386 address space limit of 3GB and dies due to random segfaults as memory allocation fails. BDB caches are set to ~1.25GB total and the entry cache to 10,000 entries. I'm not sure how long it takes slapd to consume the extra 2GB of memory, but it's definitely less than 20 minutes according to sar(1).
I have a couple core files from when this happens. Are there structures that are likely suspects for unchecked growth (the entry and BDB caches, for example) that I can poke at with a debugger to get an idea what's going on?
john