TCP stack lock contention with short-lived connections

[Julien Charbon]

 Hi,

On 7/14/16 7:38 PM, Julien Charbon wrote:
> On 6/28/16 12:06 PM, Julien Charbon wrote:
>> On 12/7/15 4:36 PM, Julien Charbon wrote:
>>> On 30/05/14 06:12, k simon wrote:
>>>>   Does any plan commit and MFC to the 10-stable ?
>>>
>>>  I got a bit of interest of having the performance improvements for
>>> short-lived TCP connections in 10-stable.  Just to share the current
>>> status to a wider audience:
>>>
>> <snip>
>>  If nobody complains, I plan to MFC them in stable/10.  After actually
>> quite a bunch of tests as I see a lot of changes that might impact these
>> MFCs (like TFO support, etc.).
> 
>  MFC over stable/10 of above commits done and tested here:
> 
> https://github.com/verisign/freebsd/commits/10/tcp-scale
> 
>  Nothing particular, the result make stable/10 much closer to -CURRENT
> TCP locking-wise.

 Only one commit remains to be pushed in stable/10:  The INP_INFO global
lock decomposition change that brings most of the performance improvement:

https://github.com/verisign/freebsd/commits/10/tcp-scale

 I am going to MFC it in stable/10.  This change is well tested in both
10 and 11 now, and will make future TCP changes MFC much easier.

 Next step for this effort:  As soon as we get a RCU-ish list
implementation in kernel for example based on:

Add ConcurrencyKit support
https://reviews.freebsd.org/D7016

 We will work on removing all usages of INP_INFO lock in TCP stack.

 Thanks.

--
Julien


-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 496 bytes
Desc: OpenPGP digital signature
URL: <http://lists.freebsd.org/pipermail/freebsd-net/attachments/20161121/8d1ca2d5/attachment.sig>