[BDW-GC] Performance impact of static allocation

[Ludovic Courtès]

Hello!

I just updated the `bdw-gc-static-alloc' branch [0].  It statically
allocates stringbufs, strings, and subrs defined using the "snarfing
macros".  It links `libguile' with `-z relro' such that constants
needing relocation are placed in a `PT_GNU_RELRO' ELF segment, which is
made read-only by the dynamic linker after relocation.

Current libgc CVS (pre-7.2) is aware of `PT_GNU_RELRO' [1, 2] and does
not scan data contained in this segment (since it cannot possibly
contain pointers to heap-allocated objects).  This was not the case in
7.1.  Thus I compared `bdw-gc-static-alloc' to `boehm-demers-weiser-gc'
using the latest libgc.  The benchmark is (again) `gcbench.scm'; the
baseline is current Guile `master':

  * no static allocation (boehm-demers-weiser-gc)

                         heap size (MiB) execution time (s.)
    Guile                   53.83 (1.00x)     21.167 (1.00x)
    BDW-GC, FSD=3           53.22 (0.99x)     15.204 (0.72x) !
    BDW-GC, FSD=6           53.11 (0.99x)     15.457 (0.73x) !
    BDW-GC, FSD=9           43.62 (0.81x)     16.810 (0.79x) !
    BDW-GC, FSD=3 incr.     63.22 (1.17x)     19.716 (0.93x)
    BDW-GC, FSD=3 gene.     96.42 (1.79x)     18.641 (0.88x)

  * static allocation, `-z relro' (bdw-gc-static-alloc)

                         heap size (MiB) execution time (s.)
    Guile                   53.83 (1.00x)     20.556 (1.00x)
    BDW-GC, FSD=3           47.33 (0.88x)     14.391 (0.70x) !
    BDW-GC, FSD=6           46.75 (0.87x)     14.914 (0.73x) !
    BDW-GC, FSD=9           44.74 (0.83x)     15.130 (0.74x) !
    BDW-GC, FSD=3 incr.     82.51 (1.53x)     19.023 (0.93x)
    BDW-GC, FSD=3 gene.     93.87 (1.74x)     17.147 (0.83x)

  * static allocation (bdw-gc-static-alloc), *without* `-z relro'

                         heap size (MiB) execution time (s.)
    Guile                   53.83 (1.00x)     21.768 (1.00x)
    BDW-GC, FSD=3           55.85 (1.04x)     16.156 (0.74x)
    BDW-GC, FSD=6           55.29 (1.03x)     15.220 (0.70x)
    BDW-GC, FSD=9           44.74 (0.83x)     16.355 (0.75x) !
    BDW-GC, FSD=3 incr.    102.09 (1.90x)     20.571 (0.95x)
    BDW-GC, FSD=3 gene.     89.70 (1.67x)     19.215 (0.88x)

    (Heap usage in incremental mode seems to be highly variable in this
    case, sometimes going as high as 128 MiB.)


Conclusions
-----------

Libgc's handling of `PT_GNU_RELRO' segments provides a noticeable
improvement.  With it, static allocation actually improves performance
(5%) and leads to reduced heap usage (11%).

However, when not linking with `-z relro', static allocation leads to
slightly degraded performance and increased heap usage (perhaps due to
misidentified pointers in the `.data.rel.ro' section?).  This is
probably worth some investigation on the BDW-GC side.

Thanks,
Ludo'.

[0] 
http://git.savannah.gnu.org/gitweb/?p=guile.git;a=commit;h=5f236208d0d864546e59afa0f5a11c9b3ba14b10
[1] 
http://thread.gmane.org/gmane.comp.programming.garbage-collection.boehmgc/2570
[2] 
http://thread.gmane.org/gmane.comp.programming.garbage-collection.boehmgc/2716

pgpdzHWJBip0y.pgp
Description: PGP signature