OK,  I found my issue.

 

The input to the hash function should include all  4 parameters also in case of symmetic:

(SRC_IP^DST_IP ,  SRC_IP^DST_IP, SRC_PORT^DST_PORT, SRC_PORT^DST_PORT)

 

Shaul.

 

From: Shaul Hamoi (shamoi) <shamoi@cisco.com>
Date: Sunday, 31 March 2024 at 14:48
To: sismis@cesnet.cz <sismis@cesnet.cz>
Cc: Balakrishnan.K1@tatacommunications.com <Balakrishnan.K1@tatacommunications.com>, stephen@networkplumber.org <stephen@networkplumber.org>, users@dpdk.org <users@dpdk.org>
Subject: Re: Symmetric RSS Hashing support in DPDK

Hi,


I am looking to determine the target queue for a flow in advance
I’m using dpdk-rss-flows.py to calculate the Toeplitz hash with I40E key.
However, If I use the script as is, I don’t obtain the correct queue because the hash calculation isn’t symmetric.
I have attempted to XOR the source and destination IP addresses  and source and destination  port before performing the calculation  but I still don’t get the correct queue.


In case you use RTE_ETH_HASH_SYMMETRIC_TOEPLITZ , the I40E key doesn’t change - Right ?
How do I find that the NIC does for symmetric hash so I can simulate it ?
(src="1.0.0.1", dst="2.0.0.2")/TCP(sport=8819, dport=80)
Actual queue - 7 (32 RX queues)
 
Result with XOR:
dpdk-rss-flows.py -s 8819 -d 80 -k i40e 32 1.0.0.1 2.0.0.2 
SRC_IP     SPORT    DST_IP     DPORT    QUEUE
1.0.0.1    8819     2.0.0.2    80       19
 
Result without (original script):
dpdk-rss-flows.py -s 8819 -d 80 -k i40e 32 1.0.0.1 2.0.0.2 
SRC_IP     SPORT    DST_IP     DPORT    QUEUE
1.0.0.1    8819     2.0.0.2    80       20
 
 
Thanks in advance,
Shaul.
 
From: "Lukáš Šišmiš" <sismis@cesnet.cz>
To: Balakrishnan K <Balakrishnan.K1@tatacommunications.com>,
        Stephen Hemminger <stephen@networkplumber.org>
Cc: "users@dpdk.org" <users@dpdk.org>
Subject: Re: Symmetric RSS Hashing support in DPDK
Date: Fri, 8 Mar 2024 08:29:32 +0100  [thread overview]
Message-ID: <90dc11cf-8c60-4b90-a3fc-43b2351b6c93@cesnet.cz> (raw)
In-Reply-To: <PSAPR04MB551616841D8221851FF79214D6272@PSAPR04MB5516.apcprd04.prod.outlook.com>
 
Hi all,
 
I've made minimalist example app on how to set symmetric RSS support for 
X710 that uses RTE_FLOW rules - check it out here:
 
https://github.com/lukashino/i40e-symmetric-rss-rte-flow
 
Lukas
 
On 08. 03. 24 6:53, Balakrishnan K wrote:
> Hi Stephen,
>    Thanks for the response . I will below option and come back if any help required.
> Regards,
> Bala
> -----Original Message-----
> From: Stephen Hemminger <stephen@networkplumber.org>
> Sent: Wednesday, March 6, 2024 8:34 PM
> To: Balakrishnan K <Balakrishnan.K1@tatacommunications.com>
> Cc: users@dpdk.org
> Subject: Re: Symmetric RSS Hashing support in DPDK
> CAUTION: This email originated from outside of the organization. Do not click links or open attachments unless you recognize the sender and know the content is safe.
> On Wed, 6 Mar 2024 07:28:40 +0000
> Balakrishnan K <Balakrishnan.K1@tatacommunications.com> wrote:
>> Hello,
>>     Our application needs symmetric hashing to handle the reverse
>> traffic on the same core, also to Improve performance by distributing the traffic across core.
>> Tried using rss config as below .
>> action_rss_tcp.types = ETH_RSS_NONFRAG_IPV4_TCP | ETH_RSS_L3_SRC_ONLY|
>> ETH_RSS_L3_DST_ONLY | ETH_RSS_L4_SRC_ONLY | ETH_RSS_L4_DST_ONLY; but could not get desired result.
>> Is there any options or API available to enable symmetric RSS hashing .
>> We are using dpdk 20.11 and intel NIC X710 10GbE .
>> 
>> Regards,
>> Bala
> With XL710 there are two choices:
>          1. Set RSS hash function to RTE_ETH_HASH_SYMMETRIC_TOEPLITZ in
>             the rte_eth_rss_conf passed in during configure
>          2. Use default (non symmetric TOEPLITZ) but pass in a rss_key that
>             has duplicated bits in the right place. Like:
> 0x6d5a 0x6d5a 0x6d5a 0x6d5a
> 0x6d5a 0x6d5a 0x6d5a 0x6d5a
> 0x6d5a 0x6d5a 0x6d5a 0x6d5a
> 0x6d5a 0x6d5a 0x6d5a 0x6d5a
> 0x6d5a 0x6d5a 0x6d5a 0x6d5a

> https://www.ndsl.kaist.edu/~kyoungsoo/papers/TR-symRSS.pdf