Re: [PATCH] Shrinks sizeof(files_struct) and better layout

[Eric Dumazet <dada1@cosmosbay.com>]

Andi Kleen a écrit :
> 
> Total data of all objects together. That's because caches always get their
> own pages and cannot share them with other caches.

OK for this part.

> The overhead of the kmem_cache_t by itself is negligible.

This seems a common misconception among kernel devs (even the best ones Andi :) )

On SMP (and/or NUMA) machines : overhead of kmem_cache_t is *big*

See enable_cpucache in mm/slab.c for 'limit' determination :

         if (cachep->objsize > 131072)
                 limit = 1;
         else if (cachep->objsize > PAGE_SIZE)
                 limit = 8;
         else if (cachep->objsize > 1024)
                 limit = 24;
         else if (cachep->objsize > 256)
                 limit = 54;
         else
                 limit = 120;

On a 64 bits machines, 120*sizeof(void*) = 120*8 = 960

So for small objects (<= 256 bytes), you end with a sizeof(array_cache) = 1024 
bytes per cpu

If 16 CPUS : 16*1024 = 16 Kbytes + all other kmem_cache structures : (If you 
have a lot of Memory Nodes, then it can be *very* big too).

If you know that no more than 100 objects are used in 99% of setups, then a 
dedicated cache is overkill, even locking 100 pages because of extreme 
fragmentation is better.

Probability that a *lot* of tasks are created at once and killed at once is 
close to 0 during a machine lifetime.


Maybe we can introduce an ultra basic memory allocator for such objects 
(without CPU caches, node caches), so that the memory overhead is small. 
Hitting a spinlock at thread creation/deletion time is not that time critical.

Eric