DragonFly BSD
DragonFly kernel List (threaded) for 2003-09
[Date Prev][Date Next]  [Thread Prev][Thread Next]  [Date Index][Thread Index]

Re: SLAB allocator now the default.

From: Sander Vesik <sander@xxxxxxxxxxxxxxxxxxx>
Date: 28 Sep 2003 23:20:00 GMT
Cache-post-path: haldjas.folklore.ee!unknown@localhost

Sander Vesik <sander@xxxxxxxxxxxxxxxxxxx> wrote:

[urkh, that was a stupid runaway post]

> Hiten Pandya <hmp@xxxxxxxxxxxxx> wrote:
>> Matthew Dillon wrote:
>>>     The slab allocator is now the default.  The NO_KMEM_MAP and
>>>     USE_SLAB_ALLOCATOR options have been removed.
>>>     USE_KMEM_MAP and NO_SLAB_ALLOCATOR options have temporarily been added,
>>>     but these too will be removed (along with the old malloc and kmem_map
>>>     code) in about a week.
>>        Matt,
>>        I haven't looked at the Slab-alloc code recently, but I am
>>        wondering if you are planning to remove the power-of-2 alignment
>>        stuff out of malloc()?
>>        From the many papers I have been reading in the last couple of
>>        weeks, it seems that power-of-2 alignment seems to cause a lot
>>        of performance degradation in SMP cases, and cache issues.

SMP? you mean the SPARCcenter example? thats more like really bad interaction
between allocations and memory interleaving than an SMP effect. But you can
probably get something similar in opteron systems if you turn interleaving on
from BIOS and possibly even from some pretty conventional dual channel chipsets
and just a single CPU.

>>        A good example of this demonstration is in Bonwick's papers.
>>        I understand this would be quite a lengthy and tedious procedure
>>        but it is something that I have been wondering.
>>        Regards,


+++ Out of cheese error +++

[Date Prev][Date Next]  [Thread Prev][Thread Next]  [Date Index][Thread Index]