-
Notifications
You must be signed in to change notification settings - Fork 46
Meeting Minutes January 4, 2018
10 am to 11:30 am, Dell-EMC, 5450 Great America Parkway, Santa Clara, CA 95054
Barefoot Networks: Jeongkeun Lee, Mickey Spiegel
Cisco Systems: Rajesh Sharma, Mario Baldi
Dell: Raja Jayakumar, Senthil Ganesan
Mellanox: Alan Lo
Netronome: Bapi Vinnakota
Netsia: Serkant
Alibaba: Jianwen Pi
Postech: Jonghwan Hyun
VMware: Mukesh Hira
Xilinx: Gordon Brebner, Robert Halstead
Slide deck from the meeting can be found here
Mickey Spiegel presented the telemetry report format draft specification. There was discussion on the following aspects of the specification
The current specification has a sub-system hw_id field in the top-level report header, but switch ID is embedded in the drop header or switch local report header. It may be better to move the switch ID to the top-level fixed header.
The report format specification currently defines three different fixed reports - a drop report, a switch local report and an inline packet report with INT metadata inside the packet. A next-protocol field is used to identify the stack of information in the report. Instead, there could be one report format with a bit mask identifying the information in the report, or a TLV format for the stack of information being reported.
The current telemetry report specification reports state for a single packet in each report. Coalescing reports for multiple packets together in a report packet would help reduce the amount of report traffic to collectors. However, this would require defining what reports may be coalesced together (e.g. reports for packets belonging to a single flow, or single queue) and defining idle timeouts upon which reports are sent to the monitor without maximum amount of coalescing.
In some use-cases, a switch may offload the telemetry report generation to a dedicated line card/processor. In that case, packets for which telemetry reports are to be generated are simply punted to this report generation engine, and the packets may need to be annotated with information to be encoded in the report (e.g. drop reason, queue state, etc). We consider the interface between the data plane and the report generation engine within a single switch is considered to be implementation-specific within a single entity and don't see a strong need for a standard format for such an interface, although the data plane and report generating engine may be developed by different vendors.
- We ran out of time and could not have a discussion on the open issues. The open issues for network telemetry can be found in the p4-applications git repository. Please elaborate on use-cases, concerns, possible solutions directly on GitHub, so we can make progress towards the first version of the specification. Also please feel free to create new issues as needed against the INT or report format specifications.
- We will hold a working group meeting once every two weeks. Next meeting will be from 10 am to 11:30 am on Thursday January 18, 2018. We will send out the agenda and meeting location information on the p4-apps mailing list.