-
Notifications
You must be signed in to change notification settings - Fork 46
Meeting Minutes: September 20, 2018
10 am to 11:30 am PDT (GMT - 7), Barefoot Networks, 4750 Patrick Henry Dr, Santa Clara, CA 95054
Barefoot Networks: Jeongkeun Lee, Mickey Spiegel
Cisco Systems: Ramesh Sivakolundu, Mario Baldi
Mellanox: Barak Gafni
VMware: Mukesh Hira
Still debating whether the proposed S bit and the action it indicates to the INT sink should be domain-specific, or part of the base INT specification. Mickey and Ramesh to get back on this, the PR will remain open until this decision is made.
We agree on the need to include support for INT over GRE encapsulation for deployments that could use it. Mickey will create a pull request with edits to the specifications for this.
Ramesh proposed the formats for carrying INT metadata in IPv6 hop by hop header, including support for deployments with IPv6 Segment Routing. Ramesh will create a pull request with relevant edits to the specification for this.
Ramesh presented a proposal for carrying INT in an MPLS extension header, allowing collection of INT metadata over a MPLS LSP. The proposal is based on use of an Extension Header Label that indicates presence of an INT extension header. This Extension Header Label is inserted by INT source and can be present anywhere in the label stack, thus requiring every INT-capable MPLS switch to parse the entire label stack to gauge the presence of INT header. Using a new Ethertype for MPLS with extension headers may be more efficient. We decided that it is early to decide on this, and will hold off on INT support over MPLS.
Mickey presented a proposal to address ECMP hashing based on User Defined Fields (UDF) in the presence of INT header. Presence of a variable length INT header between TCP/UDP header and payload would break ECMP hashing based on a UDF inserted at a certain offset from the TCP header at the traffic source. The proposal involves copying the UDF field in the packet preceding the INT headers. This will require all transit switches to be INT-aware with respect to location of UDF even if they are not all INT-capable.
Ramesh proposed defining a new INT instruction bit for buffer occupancy, allowing for queue occupancy and buffer occupancy to be reported separately in shared buffer architectures. Ramesh will create a pull request with relevant edits to the INT specifications and the YANG model for metadata semantics.
- We shall review above pull requests over GitHub and meet again in two weeks on October 4 to close on the above items.