[OpenSIPS-Devel] [OpenSIPS/opensips] fc9344: HP_MALLOC: Dramatically reduce memory fragmentation

Liviu Chircu noreply at github.com
Fri Apr 3 13:44:16 EST 2020


  Branch: refs/heads/master
  Home:   https://github.com/OpenSIPS/opensips
  Commit: fc9344a5b22d0da088bfa066cd22f2cbd2450338
      https://github.com/OpenSIPS/opensips/commit/fc9344a5b22d0da088bfa066cd22f2cbd2450338
  Author: Liviu Chircu <liviu at opensips.org>
  Date:   2020-04-03 (Fri, 03 Apr 2020)

  Changed paths:
    M mem/common.h
    M mem/hp_malloc.c
    M mem/hp_malloc.h
    M mem/hp_malloc_dyn.h
    M mem/hp_malloc_stats.c
    M mem/hp_malloc_stats.h
    M statistics.c

  Log Message:
  -----------
  HP_MALLOC:  Dramatically reduce memory fragmentation

For HP_MALLOC to offer optimal parallel allocation performance, it
needs the memory to be fragmented: the more available chunks, the better
its performance.  However, there needs to be some coalescing as well,
otherwise variable-length allocations past the 10K range will
quickly deplete (fragment) even SHM pools of 10+ GB.

This patch adds the following heuristics to HP_MALLOC and HP_MALLOC_DBG:

    * shm_malloc(): only split a fragment if the resulting chunk is at
      least MIN_SPLIT_SIZE bytes (default 256 for PKG, 4096 for SHM)
    * shm_free(): try to coalesce the next fragment (enough to fix all
      fragmentation issues)

Many thanks to 46Labs for supporting this work!


  Commit: 3634b398226306a90999318b2aeb37e99191f484
      https://github.com/OpenSIPS/opensips/commit/3634b398226306a90999318b2aeb37e99191f484
  Author: Liviu Chircu <liviu at opensips.org>
  Date:   2020-04-03 (Fri, 03 Apr 2020)

  Changed paths:
    M mem/hp_malloc_dyn.h
    M mem/hp_malloc_stats.c
    M mem/hp_malloc_stats.h

  Log Message:
  -----------
  HP_MALLOC: Fix computation of PKG stats

Many thanks to 46Labs for supporting this work!


  Commit: 0d6d33e586f050d7ea5f375cab66fc60c4dbe1a2
      https://github.com/OpenSIPS/opensips/commit/0d6d33e586f050d7ea5f375cab66fc60c4dbe1a2
  Author: Liviu Chircu <liviu at opensips.org>
  Date:   2020-04-03 (Fri, 03 Apr 2020)

  Changed paths:
    M mem/hp_malloc_dyn.h

  Log Message:
  -----------
  HP_MALLOC: Add "re-scanning" logic

Given that HP_MALLOC has fine-grained locking, if the big fragment
shifts down from hash bucket N to bucket N-1 due to another
process performing the allocation, we may actually "lose" it during
our own scan, since bucket N-1 was empty and we're now waiting
for bucket N to unlock.  When it does unlock, it will also be empty.

As a solution: retry the scan up to N times, as long as it's feasible!

Many thanks to 46Labs for supporting this work!


  Commit: 54d4af2f34fab12637c06605141fdd008b6054b9
      https://github.com/OpenSIPS/opensips/commit/54d4af2f34fab12637c06605141fdd008b6054b9
  Author: Liviu Chircu <liviu at opensips.org>
  Date:   2020-04-03 (Fri, 03 Apr 2020)

  Changed paths:
    M mem/hp_malloc_dyn.h

  Log Message:
  -----------
  HP_MALLOC: Aggregate SHM statistic updates

It is better to update the stats with a diff of the
attached/detached/split/merged fragment changes.  Updating the stats in
a "step by step" manner has the side-effect of creating temporarily
bogus values. (i.e. negative "used" and/or "real_used")

Many thanks to 46Labs for supporting this work!


Compare: https://github.com/OpenSIPS/opensips/compare/1897599fd8a5...54d4af2f34fa



More information about the Devel mailing list