Re: tipc_init(), WARNING: at arch/x86/mm/highmem_32.c:52, [2.6.24-rc4-git5: Reported regressions from 2.6.23]

[Eric Dumazet <dada1@cosmosbay.com>]

Christoph Lameter a écrit :
> On Sat, 8 Dec 2007, Ingo Molnar wrote:
> 
>>> Good. Although we should perhaps look at that reported performance 
>>> problem with SLUB. It looks like SLUB will do a memclear() for the 
>>> area twice (first for the whole page, then for the thing it allocated) 
>>> for the slow case. Maybe that exacerbates the problem.
>> i dont think the SLUB problem could be explained purely via a double 
>> memset(). [which ought to be extremely fast anyway] We are talking about 
>> a 10 times slowdown on a 64-way box of a workload that is fairly 
>> common-sense. (tasks sending messages to each other via bog standard 
>> means)
>>
>> while i dont want to jump to conclusions without looking at some 
>> profiles, i think the SLUB performance regression is indicative of the 
>> following fallacy: "SLAB can be done significantly simpler while keeping 
>> the same performance".
> 
> Well this is double crap. First of all SLUB does not do memclear twice. 
> There is no reason to assume that SLUB has the problem just because SLOB 
> hat that. A "fix" for that nonexistent problem went into Linus tree. WTH 
> is going on?
> 
> SLUB was done because of a series of problem with the basic concepts of 
> SLAB that treaten it usability in the future.
>  
>> I couldnt point to any particular aspect of SLAB that i could 
>> characterise as "needless bloat".
> 
> I agree, SLABs architecture is pretty tight and I was one of those who 
> helped it along to be that way.
> 
> However, SLAB is just fundamentally wrong for todays machine. The key 
> problem today is cacheline fetch latency and that problem will increase 
> significantly in the future. Sure under some circumstances that exploit 
> the fact that SLAB sometimes gets its guesses on the cpu cache right SLAB 
> can still win but the more processors and nodes we get the more it will 
> become difficult to keep SLAB around and the more it will become 
> difficult to establish what cachelines are in the cpu cache.
> 
>> I think we should we make SLAB the default for v2.6.24 ...
> 
> If you guarantee that all the regression of SLAB vs. SLUB are addressed 
> then thats fine but AFAICT that is not possible.
> 
> Here is a list of some of the benefits of SLUB just in case we forgot:
> 
> 
> - SLUB is performance wise much faster than SLAB. This can be more than a
>   factor of 10 (case of concurrent allocations / frees on multiple
>   processors). See http://lkml.org/lkml/2007/10/27/245
> 
> - Single threaded allocation speed is up to double that of SLAB
> 
> - Remote freeing of objectcs in a NUMA systems is typically 30% faster.
> 
> - Debugging on SLAB is difficult. Requires recompile of the kernel
>   and the resulting output is difficult to interpret. SLUB can apply
>   debugging options to a subset of the slabcaches in order to allow
>   the system to work with maximum speed. This is necessary to detect
>   difficult to reproduce race conditions.
> 
> - SLAB can capture huge amounts of memory in its queues. The problem
>   gets worse the more processors and NUMA nodes are in the system. The 
>   amount of memory limits the number of per cpu objects one can configure.
> 
> - SLAB requires a pass through all slab caches every 2 seconds to
>   expire objects. This is a problem both for realtime and MPI jobs
>   that cannot take such a processor outage.
> 
> - SLAB does not have a sophisticated slabinfo tool to report the
>   state of slab objects on the system. Can provide details of
>   object use.
> 
> - SLAB requires the update of two words for freeing
>   and allocation. SLUB can do that by updating a single
>   word which allows to avoid enabling and disabling interrupts if
>   the processor supports an atomic instruction for that purpose.
>   This is important for realtime kernels where special measures
>   may have to be implemented if one wants to disable interrupts.
> 
> - SLAB requires memory to be set aside for queues (processors
>   times number of slabs times queue size). SLUB requires none of that.
> 
> - SLUB merges slab caches with similar characteristics to
>   reduce the memory footprint even further.
> 
> - SLAB performs object level NUMA management which creates
>   a complex allocator complexity. SLUB manages NUMA on the level of
>   slab pages reducing object management overhead.
> 
> - SLUB allows remote node defragmentation to avoid the buildup
>   of large partial lists on a single node.
> 
> - SLUB can actively reduce the fragmentation of slabs through
>   slab cache specific callbacks (not merged yet)
> 
> - SLUB has resiliency features that allow it to isolate a problem
>   object and continue after diagnostics have been performed.
> 
> - SLUB creates rarely used DMA caches on demand instead of creating
>   them all on bootup (SLAB).
> 

Yes, SLUB should be the way to go, but some issues are not yet solved.

I had to switch back to SLAB on a production NUMA server, with 2 nodes and 8GB 
ram. Using a lot of sockets, so a large part of memory was used by kernel.

SLUB kernel was hitting OOM after 2 or 3 days of uptime.
SLAB kernel never hit this.

Unfortunatly I dont have a test machine to reproduce the setup.

Maybe the problem is not related to SLUB at all, but an underlying VM/NUMA bug.

The /proc/buddyinfo showed that :

Node 0 contained two zones (DMA and DMA32) total 4 GB
Node 1 contained one zone (Normal) total 4 GB

So Node 0 contained no (Normal) zone

part of /proc/meminfo

Slab:          3338512 kB
SReclaimable:   789716 kB
SUnreclaim:    2548796 kB

I remember network interrupts were taken by CPU 1, so most allocations were 
done by CPU 1 (node 1), and many freeing were done on CPU 0

Hope this helps