Static Member Selection Modes For ECMP and LAG

[dkadia-marvell]

This proposal introduces the below two types of light weight static member selection modes for ECMP/LAG.

• Random
• Round Robin

[rck-innovium]

@JaiOCP @tjchadaga

Summary of the discussion on 11/14/2024

1. [Jai] The option to enable packet spraying is in practice needed at the Switch level and not at a per Next Hop Group level.
2. [Jai] A possible interpretation of SAI_SWITCH_ATTR_ECMP_DEFAULT_HASH_ALGORITHM:SAI_HASH_ALGORITHM_RANDOM is to enable packet spraying at a switch level.
   [Ravi] The only description available in the community is: /** Random-based hash algorithm */. This suggests it is still a hash algorithm i.e., it hashes on some fields of the packet. A possible example of such an algorithm could be: https://patents.google.com/patent/US8503456

Next steps agreed:

[Ravi] and [Jai] to see if there are use-cases for enabling packet-spraying at a per NHG granularity instead of switch level. We should also see what happens to telemetry packets, control packets etc. We will discuss offline over the PR and if needed use a follow up meeting.

[tjchadaga]

@dkadia-marvell, @rck-innovium - any conclusion on this? does this need another discussion?

[kcudnik]

place this between 6 and 8

[rck-innovium]

Thanks @JaiOCP for the discussion. Please find the summary below:

Summary of the follow-up discussion on 02/27/2025

• Offline Discussion with SAI Vendors: The consensus was to use SAI_HASH_ALGORITHM_RANDOM for per-packet random spraying. We have expanded the existing comments to ensure this is unambiguous.

• Hash Algorithm Selection: We agreed on the need to select the SAI hash algorithm at either the LAG level or the Nexthop Group (NHG) level. The use case presented involved a TOR switch connected to multiple appliances, each via a LAG. Some appliances can handle packet reordering, while others cannot. This approach can also be extended to NHG if the appliance is connected via ECMP.

• Concerns Raised by Jai: Jai raised a concern about exposing the hash algorithm at the NHG/LAG level. In some devices, this could lead to running out of hash generation blocks (e.g., some NHGs use CRC, some use XOR, some use CRC_XOR, etc.). We agreed to address this concern in a separate proposal to discover capacity/CRM.

@JaiOCP @tjchadaga

Summary of the discussion on 11/14/2024

1. [Jai] The option to enable packet spraying is in practice needed at the Switch level and not at a per Next Hop Group level.
2. [Jai] A possible interpretation of SAI_SWITCH_ATTR_ECMP_DEFAULT_HASH_ALGORITHM:SAI_HASH_ALGORITHM_RANDOM is to enable packet spraying at a switch level.
   [Ravi] The only description available in the community is: /** Random-based hash algorithm */. This suggests it is still a hash algorithm i.e., it hashes on some fields of the packet. A possible example of such an algorithm could be: https://patents.google.com/patent/US8503456

Next steps agreed:

[Ravi] and [Jai] to see if there are use-cases for enabling packet-spraying at a per NHG granularity instead of switch level. We should also see what happens to telemetry packets, control packets etc. We will discuss offline over the PR and if needed use a follow up meeting.

[kcudnik]

maybe we should resign with setting this value explicitly, and just leave it as last attribute to not have to change that explicitly every time