SLUB vs SLAB allocator with respect to 3.x and 4.x kernels

SLUB vs SLAB allocator with respect to 3.x and 4.x kernels

[Navin Parakkal]

Hi,

 I  found that in many worst case scenarios like fragmention of
allocator , slub performs well than slab.
  I also noticed that Centos /Ubuntu etc switched to SLUB but SLES
still uses SLAB in the default image.

Any particular reason where SLAB is the choice ?


I ask this since i find SLUB to be the default choice since 2.6.23 as
per https://en.wikipedia.org/wiki/SLUB_(software)



Regards,
Navin

Re: SLUB vs SLAB allocator with respect to 3.x and 4.x kernels

[Austin S Hemmelgarn]

[-- Attachment #1: Type: text/plain, Size: 882 bytes --]

On 2015-08-27 10:26, Navin Parakkal wrote:
> Hi,
>
>   I  found that in many worst case scenarios like fragmention of
> allocator , slub performs well than slab.
>    I also noticed that Centos /Ubuntu etc switched to SLUB but SLES
> still uses SLAB in the default image.
>
> Any particular reason where SLAB is the choice ?
>
>
> I ask this since i find SLUB to be the default choice since 2.6.23 as
> per https://en.wikipedia.org/wiki/SLUB_(software)
>
IIRC, SLUB isn't available/reliable on some arches (that and in some 
setups, the worst case scenarios are likely to never happen (for 
example, a machine with 1T of RAM that get's rebooted daily isn't likely 
to have significant issues with memory fragmentation)).

SLES probably hasn't switched because they're an enterprise distro, 
which in turn means changing as little as possible when updating.



[-- Attachment #2: S/MIME Cryptographic Signature --]
[-- Type: application/pkcs7-signature, Size: 3019 bytes --]

Re: SLUB vs SLAB allocator with respect to 3.x and 4.x kernels

[David Rientjes]

On Thu, 27 Aug 2015, Navin Parakkal wrote:

> Hi,
> 
>  I  found that in many worst case scenarios like fragmention of
> allocator , slub performs well than slab.
>   I also noticed that Centos /Ubuntu etc switched to SLUB but SLES
> still uses SLAB in the default image.
> 
> Any particular reason where SLAB is the choice ?
> 

Slab doesn't have a reliance on high-order allocations for performance 
where fragmentation is a problem, it can use a smaller footprint due to 
slub's per-cpu partial slabs, it is faster on some networking round-robin 
benchmarks on nUMA machines, and it is has less impact when implementing 
full kmem accounting for memcg.