Re: [PATCH RFC 0/2] Optimize the snmp stat aggregation for large cpus

[Raghavendra K T <raghavendra.kt@linux.vnet.ibm.com>]

On 08/26/2015 07:39 PM, Eric Dumazet wrote:
> On Wed, 2015-08-26 at 15:55 +0530, Raghavendra K T wrote:
>> On 08/26/2015 04:37 AM, David Miller wrote:
>>> From: Raghavendra K T <raghavendra.kt@linux.vnet.ibm.com>
>>> Date: Tue, 25 Aug 2015 13:24:24 +0530
>>>
>>>> Please let me know if you have suggestions/comments.
>>>
>>> Like Eric Dumazet said the idea is good but needs some adjustments.
>>>
>>> You might want to see whether a per-cpu work buffer works for this.
>>
>> sure, Let me know if I understood correctly,
>>
>> we allocate the temp buffer,
>> we will have a  "add_this_cpu_data" function and do
>>
>> for_each_online_cpu(cpu)
>>                   smp_call_function_single(cpu, add_this_cpu_data, buffer, 1)
>>
>> if not could you please point to an example you had in mind.
>
>
> Sorry I do not think it is a good idea.
>
> Sending an IPI is way more expensive and intrusive than reading 4 or 5
> cache lines from memory (per cpu)
>
> Definitely not something we want.

Okay. Another problem I thought here was that we could only loop over
online cpus.

>>> It's extremely unfortunately that we can't depend upon the destination
>>> buffer being properly aligned, because we wouldn't need a temporary
>>> scratch area if it were aligned properly.
>>
>> True, But I think for 64 bit cpus when (pad == 0) we can go ahead and
>> use stats array directly and get rid of put_unaligned(). is it correct?
>
>
> Nope. We have no alignment guarantee. It could be 0x............04
> pointer value. (ie not a multiple of 8)
>
>>
>> (my internal initial patch had this version but thought it is ugly to
>> have ifdef BITS_PER_LONG==64)
>
> This has nothing to do with arch having 64bit per long. It is about
> alignment of a u64.
>

Okay. I 'll send V2 with declaring tmp buffer in stack.