[dpdk-dev] [PATCH v2 0/3] Predictable RSS feature

[Medvedkin, Vladimir]

Hi Yipeng,

Thanks for the review,

On 10/04/2021 03:32, Wang, Yipeng1 wrote:
>> -----Original Message-----
>> From: Medvedkin, Vladimir <vladimir.medvedkin at intel.com>
>> Sent: Tuesday, April 6, 2021 12:51 PM
>> To: dev at dpdk.org
>> Cc: Ananyev, Konstantin <konstantin.ananyev at intel.com>; Chilikin, Andrey
>> <andrey.chilikin at intel.com>; Kinsella, Ray <ray.kinsella at intel.com>; Wang,
>> Yipeng1 <yipeng1.wang at intel.com>; Gobriel, Sameh
>> <sameh.gobriel at intel.com>; Richardson, Bruce
>> <bruce.richardson at intel.com>
>> Subject: [PATCH v2 0/3] Predictable RSS feature
>>
>> This patch series introduces predictable RSS feature.
>> It is based on the idea of searching for partial hash collisions within Toeplitz
>> hash.
>>
>> The Toeplitz hash function is a homomorphism between (G, ^) and (H, ^),
>> where (G, ^) - is a group of tuples and (H, ^) is a group of hashes with respect
>> to XOR operation. So tuples and hashes could be treated as n-dimension and
>> 32-dimension vector spaces over GF(2).
>> So, f(x ^ y) == f(x) ^ f(y)
>> where f - is the toeplitz hash function and x, y are tuples.
>>
>> The ability to predict partial collisions allows user to compute input hash value
>> with desired LSB values.
>> Usually number of LSB's are defined by the size of RSS Redirection Table.
>>
>> There could be number of use cases, for example:
>> 1) NAT. Using this library it is possible to select a new port number on a
>> translation in the way that rss hash for original tuple will have the same LSB's
>> as rss hash for reverse tuple.
>> 2) IPSec/MPLS/Vxlan. It is possible to choose tunnel id to be pinned to a
>> desired queue.
>> 3) TCP stack. It is possible to choose a source port number for outgoing
>> connections in the way that received replies will be assigned to desired
>> queue.
>> 4) RSS hash key generation. Hash key initialization with random values does
>> not guarantee an uniform distribution amongst queues. This library uses
>> mathematically proved algorithm to complete the rss hash key to provide the
>> best distribution.
>>
>> v2:
>> - added extra API rte_thash_adjust_tuple()
>> - added extra tests for rte_thash_adjust_tuple()
>> - added extra fields to rte_thash_subtuple_helper struct
>> - fixed typos
>>
>> Vladimir Medvedkin (3):
>>    hash: add predictable RSS API
>>    hash: add predictable RSS implementation
>>    test/hash: add additional thash tests
>>
>>   app/test/test_thash.c       | 468 +++++++++++++++++++++++++++++++-
>>   lib/librte_hash/meson.build |   3 +-
>>   lib/librte_hash/rte_thash.c | 637
>> ++++++++++++++++++++++++++++++++++++++++++++
>>   lib/librte_hash/rte_thash.h | 180 +++++++++++++
>>   lib/librte_hash/version.map |   8 +
>>   5 files changed, 1289 insertions(+), 7 deletions(-)  create mode 100644
>> lib/librte_hash/rte_thash.c
>>
>> --
>> 2.7.4
> 
> [Wang, Yipeng]
> Hi, Vladimir, thanks for the patch!
> I haven't fully understood every bit of the algorithm yet,
> but I did see issues that this patch could potentially solve.
> My understanding is that there are some restrictions for the current implementation,
> for example, it only supports port(16-bit) manipulation, but not multiple fields or IP.

It supports any bit-range ( with size >= reta_sz, configured for the 
CTX) manipulations within a tuple as well as multiple number of them.

> Still, I think it should be good for the use cases you listed. I would love to hear
> more feedbacks from people who are more familiar with doing NAT in production systems.
> 
> For me, besides the comments I sent earlier,
> good documentation and references are needed with clear usage examples, as others pointed out already.
> 
> Also, the current API design seems a bit cumbersome.
> To use the library, one needs:
> Init_ctx
> Add_helper.
> Get_helper
> Get_complement
> Then in a loop:
> Adjust_tuples
> Then XOR with the current tuple
> 
> I wonder if an alternative all-in-one API could be designed for simpler use cases.

Agree, v3 will have a new rte_thash_adjust_tuple() implementation that 
will make everything much easier.

> 
> Thanks!
> >
> 

-- 
Regards,
Vladimir