From mboxrd@z Thu Jan  1 00:00:00 1970
From: compudj@krystal.dyndns.org (Mathieu Desnoyers)
Date: Tue, 17 Aug 2010 16:51:44 -0400
Subject: [ltt-dev] liburcu cache line size
In-Reply-To: <4C6AF3A2.6000107@polymtl.ca>
References: <4C6AD54F.4080006@polymtl.ca> <20100817185122.GA4384@Krystal>
	<4C6ADDAA.6020408@polymtl.ca> <20100817194501.GA19351@Krystal>
	<4C6AE9F1.2090300@polymtl.ca> <20100817202430.GA22705@Krystal>
	<4C6AF3A2.6000107@polymtl.ca>
Message-ID: <20100817205144.GA25723@Krystal>

* David Goulet (david.goulet at polymtl.ca) wrote:
>
>
> On 10-08-17 04:24 PM, Mathieu Desnoyers wrote:
>> * David Goulet (david.goulet at polymtl.ca) wrote:
>>> On 10-08-17 03:45 PM, Mathieu Desnoyers wrote:
>> [...]
>>>> Yes. The performance degradation caused by cache-line bouncing is _way_
>>>> worse than extra cache pressure.
>>>>
>>>
>>> There is something I don't understand here. Correct me if (most likely)
>>> I am wrong.
>>>
>>> How cache line bouncing is affected by the cache line size? If I
>>> understand correctly, cache line bounce is the problem where CPUs shares
>>> data and have to fetch it from CPU0 to CPU7 (between caches). And, I
>>> surely agree, this is costly!
>>
>> That's ok up to here.
>>
>>>
>>> However, if the size of the cache is bigger then the normal cache, you
>>> just loose space... For arch with 64 cache line size, you loose two line
>>> per structure aligned... How lowering down to 64 bytes will cause cache
>>> line bouncing?
>>
>> Let's take the following example:
>>
>> A multiprocessor machine with 256 bytes cache line size.
>> The program is built thinking the cache line size is only 128 bytes.
>>
>> So we allocate an array of what we hope are per-cpu variables:
>>
>>   malloc(nr_cpus * sizeof(struct type));
>>
>> Where struct type is __attribute__((aligned(128))
>>
>> So we end up having two structures sharing a cache-line, and these will
>> bounce between CPUs, even though the structures are not shared: only the
>> cache-lines are shared, because the structures happen to be on the same
>> cache line.
>>
>> So for allocation of individual objects which are meant to be per-cpu,
>> e.g. a structure controlling the per-cpu buffer, the allocator can put
>> one structure next to another (belonging to another cpu), thus causing
>> cache line bouncing.
>>
>> This phenomenon is called "false sharing".
>>
>
> Very nice. That clarify yes!
>
> However, please refer to Intel? 64 and IA-32 Architectures Software  
> Developer's Manual Volume 3A: System Programming Guide.
>
> http://www.intel.com/Assets/PDF/manual/253668.pdf
>
> P. 527, Table 11-1
>
> ? Pentium 4 and Intel Xeon processors (Based on Intel NetBurst
>  microarchitecture): 8-KByte, 4-way set associative, 64-byte cache line
> size.
> ? Pentium 4 and Intel Xeon processors (Based on Intel NetBurst
>  microarchitecture): 16-KByte, 8-way set associative, 64-byte cache line
> size.

Dunno why the Linux kernel choses that for P4. But we definitely have to
handle NUMA systems.

Mathieu

>
> David
>
>> Mathieu
>>
>>>
>>> Thanks for your help on that!
>>> David
>>>
>>
>
> -- 
> David Goulet
> LTTng project, DORSAL Lab.
>
> PGP/GPG : 1024D/16BD8563
> BE3C 672B 9331 9796 291A  14C6 4AF7 C14B 16BD 8563
>

-- 
Mathieu Desnoyers
Operating System Efficiency R&D Consultant
EfficiOS Inc.
http://www.efficios.com