-
Notifications
You must be signed in to change notification settings - Fork 46
Meeting Minutes: March 1, 2018
10 am to 11:30 am, VMware, 3425 Hillview Avenue, Palo Alto CA 94304
Barefoot Networks - Jeongkeun Lee, Mickey Spiegel
Cisco Systems: Mario Baldi
Dell : Anoop Ghanwani, Senthil Ganesan
Mellanox : Alan Lo
Netronome: Bapi Vinnakota
Postech: Jonghwan Hyun
UFRGS: Jonatas Marques
VMware - Mukesh Hira
No slides were presented at this meeting. The discussion focused on closing on the following open issues -
JK Lee presented proposed edits to allow use of DSCP bits for indicating presence of INT in native TCP/UDP traffic, without reserving a specific bit in the DSCP field. The edits also include a definition of INT domain as it is relevant to this topic. The group agreed with the edits, JK will push these changes to the spec.
The group agreed to add checksum complement metadata field. This should be the last metadata inserted at each hop. INT source indicates whether each transit hop adds checksum complement metadata. Two options were evaluated for this indication -
(1) Using the last bit in the current 16-bit instruction bit mask so that checksum complement remains the last metadata even as we use more reserved instruction bits in the future. If we expand the instruction bit mask beyond 16 bits in a certain version, that would break backward compatibility anyway and at that point we can relocate the checksum complement bit to the new last instruction bit.
(2) Using a flag outside the instruction bit mask, each transit hop inspects this flag and if it is set, inserts checksum complement metadata at the end of the metadata stack. This makes checksum complement option independent of the instruction bit mask ensuring that checksum complement will always be the last metadata regardless of how the instruction bit mask changes over versions.
The group preferred option 1, as this indication is inherently an instruction bit. Also, a transit hop should ideally parse a contiguous set of bits for the metadata to be inserted.
Couple of questions were raised -
(1) Should we reserve a 32-bit instruction bit mask and use the last of 32 bits for checksum complement metadata to reduce the likelihood that we will run out of instruction bit space in the foreseeable future.
(2) Should the INT source also be able to indicate to transit hops that Layer 4 checksum is not being used at all and that transit hops neither need to update the Layer 4 checksum nor compute a checksum complement metadata value (e.g. when INT source is a VXLAN tunnel endpoint). Could INT transit hops simply infer this by seeing a zero Layer 4 checksum?
Based on discussion in the Feb 15 meeting, Mukesh presented edits for adding a second level of ingress port and egress port identifiers in INT metadata, each being a full 4-Byte metadata field to allow for adequately large number of logical ports per network device. There was a suggestion to use only one bit in the instruction bit mask even though this corresponds to two metadata fields. The “Instruction Count” field will now reflect the number of 4-Byte metadata fields inserted at each node, not the number of bits set in the instruction bit mask. Accordingly, this field needs to be renamed to “Node Metadata Length”.
Action item: Mukesh to incorporate above comments and amend pull request.
- Next bi-weekly meeting will be on Thursday March 15, 2018, 10 am to 11:30 am. Meeting location and agenda will be announced closer to the meeting date.
- Next meeting is the last chance to raise any issues to be addressed in version 1.0 of the specifications. We want to allow a two week buffer for making editorial changes to the specifications and publish version 1.0 at the March 29th working group meeting.
- As we approach version 1.0 of Network Telemetry, we are inviting proposals from working group members for presentations on their INT use-cases or other P4 applications of interest.