[<prev] [next>] [thread-next>] [day] [month] [year] [list]
Message-Id: <20070311021009.19963.11893.sendpatchset@schroedinger.engr.sgi.com>
Date: Sat, 10 Mar 2007 18:10:09 -0800 (PST)
From: Christoph Lameter <clameter@....com>
To: akpm@...l.org
Cc: linux-mm@...ck.org, linux-kernel@...r.kernel.org,
Christoph Lameter <clameter@....com>, mpm@...enic.com
Subject: [SLUB 0/3] SLUB: The unqueued slab allocator V5
[PATCH] SLUB The unqueued slab allocator v4
V4->V5:
- Single object slabs only for slabs > slub_max_order otherwise generate
sufficient objects to avoid frequent use of the page allocator. This is
necessary to compensate for fragmentation caused by frequent uses of
the page allocator. We expect slabs of PAGE_SIZE from this rule since
multi object slabs require uses of fields that are in use on i386 and
x86_64. See the quicklist patchset for a way to fix that issue
and a patch to get rid of the PAGE_SIZE special casing.
- Drop pass through to page allocator due to page allocator fragmenting
memory. The buffering through large order allocations is done in SLUB.
Infrequent larger order allocations cause less fragmentation
than frequent small order allocations.
- We need to update object sizes when merging slabs otherwise kzalloc
will not initialize the full object (this caused the failure on
varios platforms).
- Padding checks before redzone checks so that we get messages about
the corruption of whole slab and not about a single object.
Note that SLUB will warn on zero sized allocations. SLAB just allocates
some memory. So some traces from the usb subsystem etc should be expected.
Note that the definition of the return type of ksize() is currently
different between mm and Linus tree. Patch is conforming to mm.
V3->V4
- Rename /proc/slabinfo to /proc/slubinfo. We have a different format after
all.
- More bug fixes and stabilization of diagnostic functions. This seems
to be finally something that works wherever we test it.
- Serialize kmem_cache_create and kmem_cache_destroy via slub_lock (Adrian's
idea)
- Add two new modifications (separate patches) to guarantee
a mininum number of objects per slab and to pass through large
allocations.
V2->V3
- Debugging and diagnostic support. This is runtime enabled and not compile
time enabled. Runtime debugging can be controlled via kernel boot options
on an individual slab cache basis or globally.
- Slab Trace support (For individual slab caches).
- Resiliency support: If basic sanity checks are enabled (via F f.e.)
(boot option) then SLUB will do the best to perform diagnostics and
then continue (i.e. mark corrupted objects as used).
- Fix up numerous issues including clash of SLUBs use of page
flags with i386 arch use for pmd and pgds (which are managed
as slab caches, sigh).
- Dynamic per CPU array sizing.
- Explain SLUB slabcache flags
V1->V2
- Fix up various issues. Tested on i386 UP, X86_64 SMP, ia64 NUMA.
- Provide NUMA support by splitting partial lists per node.
- Better Slab cache merge support (now at around 50% of slabs)
- List slab cache aliases if slab caches are merged.
- Updated descriptions /proc/slabinfo output
This is a new slab allocator which was motivated by the complexity of the
existing code in mm/slab.c. It attempts to address a variety of concerns
with the existing implementation.
A. Management of object queues
A particular concern was the complex management of the numerous object
queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for
each allocating CPU and use objects from a slab directly instead of
queueing them up.
B. Storage overhead of object queues
SLAB Object queues exist per node, per CPU. The alien cache queue even
has a queue array that contain a queue for each processor on each
node. For very large systems the number of queues and the number of
objects that may be caught in those queues grows exponentially. On our
systems with 1k nodes / processors we have several gigabytes just tied up
for storing references to objects for those queues This does not include
the objects that could be on those queues. One fears that the whole
memory of the machine could one day be consumed by those queues.
C. SLAB meta data overhead
SLAB has overhead at the beginning of each slab. This means that data
cannot be naturally aligned at the beginning of a slab block. SLUB keeps
all meta data in the corresponding page_struct. Objects can be naturally
aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte
boundaries and can fit tightly into a 4k page with no bytes left over.
SLAB cannot do this.
D. SLAB has a complex cache reaper
SLUB does not need a cache reaper for UP systems. On SMP systems
the per CPU slab may be pushed back into partial list but that
operation is simple and does not require an iteration over a list
of objects. SLAB expires per CPU, shared and alien object queues
during cache reaping which may cause strange hold offs.
E. SLAB has complex NUMA policy layer support
SLUB pushes NUMA policy handling into the page allocator. This means that
allocation is coarser (SLUB does interleave on a page level) but that
situation was also present before 2.6.13. SLABs application of
policies to individual slab objects allocated in SLAB is
certainly a performance concern due to the frequent references to
memory policies which may lead a sequence of objects to come from
one node after another. SLUB will get a slab full of objects
from one node and then will switch to the next.
F. Reduction of the size of partial slab lists
SLAB has per node partial lists. This means that over time a large
number of partial slabs may accumulate on those lists. These can
only be reused if allocator occur on specific nodes. SLUB has a global
pool of partial slabs and will consume slabs from that pool to
decrease fragmentation.
G. Tunables
SLAB has sophisticated tuning abilities for each slab cache. One can
manipulate the queue sizes in detail. However, filling the queues still
requires the uses of the spin lock to check out slabs. SLUB has a global
parameter (min_slab_order) for tuning. Increasing the minimum slab
order can decrease the locking overhead. The bigger the slab order the
less motions of pages between per CPU and partial lists occur and the
better SLUB will be scaling.
G. Slab merging
We often have slab caches with similar parameters. SLUB detects those
on boot up and merges them into the corresponding general caches. This
leads to more effective memory use. About 50% of all caches can
be eliminated through slab merging. This will also decrease
slab fragmentation because partial allocated slabs can be filled
up again. Slab merging can be switched off by specifying
slub_nomerge on boot up.
Note that merging can expose heretofore unknown bugs in the kernel
because corrupted objects may now be placed differently and corrupt
differing neighboring objects. Enable sanity checks to find those.
H. Diagnostics
The current slab diagnostics are difficult to use and require a
recompilation of the kernel. SLUB contains debugging code that
is always available (but is kept out of the hot code paths).
SLUB diagnostics can be enabled via the "slab_debug" option.
Parameters can be specified to select a single or a group of
slab caches for diagnostics. This means that the system is running
with the usual performance and it is much more likely that
race conditions can be reproduced.
I. Resiliency
If basic sanity checks are on then SLUB is capable of detecting
common error conditions and recover as best as possible to allow the
system to continue.
J. Tracing
Tracing can be enabled via the slab_debug=T,<slabcache> option
during boot. SLUB will then protocol each action on that slabcache
and dump the object contents on free.
K. On demand DMA cache creation.
Generally DMA caches are not needed. If a kmalloc is used with
__GFP_DMA then just create this single slabcache that is needed.
For systems that have no ZONE_DMA requirement the support is
completely eliminated.
L. Performance increase
Some benchmarks have shown speed improvements on kernbench in the
range of 5-10%. The locking overhead of slub is based on the
underlying base allocation size. If we can reliably allocate
larger order pages then it is possible to increase slub
performance much further. The anti-fragmentation patches may
enable further performance increases.
Tested on:
i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator
SLUB Boot options
slub_nomerge Disable merging of slabs
slub_min_order=x Require a minimum order for slab caches. This
increases the managed chunk size and therefore
reduces meta data and locking overhead.
slub_min_objects=x Mininum objects per slab. Default is 8.
slub_max_order=x Avoid generating slab large than order specified.
slub_debug Enable all diagnostics for all caches
slub_debug=<options> Enable selective options for all caches
slub_debug=<o>,<cache> Enable selective options for a certain set of
caches
Available Debug options
F Double Free checking, sanity and resiliency
R Red zoning
P Object / padding poisoning
U Track last free / alloc
T Trace all allocs / frees (only use on individual slabs).
To use SLUB: Apply this patch and then select SLUB as the default slab
allocator. The output of /proc/slabinfo will then change. Here is a
sample (this is an UP/SMP format. The NUMA display will show on which nodes
the slabs were allocated). Flags are
a Cpucache Align requested
A Hardware Align required
C Constructor
d DMA cache
D Destructor
F Double free checking/Sanity
p Panic on failure
P Poisoning
r Objects are reclaimable
R RCU destroy
S Memory Spreading
U User Tracking
T Tracing
Z Red Zone
Thanks to Adrian Drzewiecki <z@...e.net> for many ideas and spotting a lot of
bugs.
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
Powered by blists - more mailing lists