> There is updated graphs on
> https://github.com/arneg/GJAlloc/wiki/Performance-comparison. Turns
> out, in fact, I misattributed benefits to the std::allocator it does
> not have.
Nice, thank you. Afaik there are no near-term 7.9 release plans, so
I'm for letting this allocator go in, provided there's a configure
option to switch to the old one.
I'd rather have it directly in the source than as a bundle, but I
guess it's up to you really. (I tried to symlink the bundle dir to
your upstream git, but I couldn't get that to work very well - clearly
the bundle system isn't made with that in mind. But that's fixable, of
course.)
> And my bad again. Currently the code does not do this, for
> comparability between the two block allocators as used by CritBits,
> I think.
Do you have any plans for switching to object-local storage? I'm not
sure a single shared pool would do that well if there are many objects
- I suspect it'd be bad for locality, and if we ever go properly
multi-cpu it'd produce a lot of false sharing in cpu caches.
When I implemented the new multisets based on red-black trees, the
node memory management was actually the most time consuming part. My
implementation separates the nodes for each multiset, and moves them
to be able to resize the memory blocks, and to compact them to limit
internal fragmentation. It always allocates a single block for the
whole multiset, regardless how large. It'd probably be better to use
multiple blocks when it fills more than one page, but it could still
be important to support moving nodes to be able to compact smaller
trees.
Note that handling moving nodes has big effect on the iterators. They
need to address nodes by index rather than by direct pointer, if they
should be able to handle simultaneous change in the tree. For
multisets there's a "semi-frozen" state when there are active
iterators: Since iterators address by index, the memory block as a
whole may move (and thus be enlarged), but the nodes inside it may
not, so compaction is postponed until all iterators are gone. (There's
a comment near the top of multiset.h that goes into detail on this.)
So my point is just that handling object-local storage of the nodes
could have a quite big effect on the code.
> /.../ any chance, your src/post_modules/CritBit/dependencies has
> size > 0?
Hmm, I'm pretty sure I tried "make depend" a few times before giving
up yesterday, but now it works. Sorry for the noise.
> > And of course, a couple of CritBit benchmarks would be welcome - they
> > also ought to show off GJAlloc, right? ;)
>
> That can be arranged.
Cool. I've added a test that shows the difference better: It allocates
and frees random objects while keeping 1000000 objects allocated. That
shows a clear difference:
total user mem (runs)
Old BLOCK_ALLOC 2.036s 1.689s 89112kb (3) (0/s)
GJAlloc 1.110s 0.805s 89084kb (5) (0/s)
Interesting for Roxen as well, since Roxen servers tend to have quite
a lot of objects allocated. Sweet. :)
