PATCH: two slab-like memory allocators - Mailing list pgsql-hackers

From Tomas Vondra
Subject PATCH: two slab-like memory allocators
Date
Msg-id d15dff83-0b37-28ed-0809-95a5cc7292ad@2ndquadrant.com
Whole thread Raw
Responses Re: PATCH: two slab-like memory allocators  (Petr Jelinek <petr@2ndquadrant.com>)
[HACKERS] PATCH : Generational memory allocator (was PATCH: two slab-likememory allocators)  (Tomas Vondra <tomas.vondra@2ndquadrant.com>)
List pgsql-hackers
Hi,

Back in the bug #14231 thread [1], dealing with performance issues in
reorderbuffer due to excessive number of expensive free() calls, I've
proposed to resolve that by a custom slab-like memory allocator,
suitable for fixed-size allocations. I'd like to put this into the next
CF, and it's probably too invasive change to count as a bugfix anyway.

[1]
https://www.postgresql.org/message-id/flat/20160706185502.1426.28143%40wrigleys.postgresql.org

This patch actually includes two new memory allocators (not one). Very
brief summary (for more detailed explanation of the ideas, see comments
at the beginning of slab.c and genslab.c):

Slab
----
* suitable for fixed-length allocations (other pallocs fail)
* much simpler than AllocSet (no global freelist management etc.)
* free space is tracked per block (using a simple bitmap)
* which allows freeing the block once all chunks are freed (AllocSet
will hold the memory forever, in the hope of reusing it)

GenSlab
-------
* suitable for non-fixed-length allocations, but with chunks of mostly
the same size (initially unknown, the context will tune itself)
* a combination AllocSet and Slab (or a sequence of Slab allocators)
* the goal is to do most allocations in Slab context
* there's always a single 'current' Slab context, and every now and and
then it's replaced with a new generation (with the chunk size computed
from recent requests)
* the AllocSet context is used for chunks too large for current Slab

So none of this is meant as a universal replacement of AllocSet, but in
the suitable cases the results seem really promising. For example for
the simple test query in [1], the performance improvement is this:

         N |   master |  patched
    -----------------------------
     10000 |    100ms |    100ms
     50000 |  15000ms |    350ms
    100000 | 146000ms |    700ms
    200000 |        ? |   1400ms

That's a fairly significant improvement, and the submitted version of
the patches should perform even better (~2x, IIRC).

There's a bunch of TODOs - e.g. handling of realloc() calls in the
GenSlab, and probably things I haven't thought about.

regards

--
Tomas Vondra                  http://www.2ndQuadrant.com
PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services

Attachment

pgsql-hackers by date:

Previous
From: Simon Riggs
Date:
Subject: Re: Why we lost Uber as a user
Next
From: Andrew Gierth
Date:
Subject: Re: Wanting to learn about pgsql design decision