-
Notifications
You must be signed in to change notification settings - Fork 46
Meeting Minutes: May 17, 2018
10 am to 11:30 am PDT (GMT - 7), Dell-EMC, 5450 Great America Parkway, Santa Clara, CA 95054
AT&T: Shyam Parekh, Tom Tofigh
Barefoot Networks: Jeongkeun Lee, Mickey Spiegel
Cisco Systems: Mario Baldi
Dell: Senthil Ganesan, Raja Jayakumar
Intel: Michael Orr
IXIA: Chris Sommers
Juniper Networks: Sandesh Kumar Sodhi
Mellanox: Barak Gafni
Netsia: Serkant Uluderya
Postech: Jonghwan Hyun
Surfnet: Ronald Vanderpol
VMware: Mukesh Hira
Tom Tofigh and Shyam Parekh from AT&T presented new "Time Shifting" use-cases based on using in-network storage (memory/SSD) to store all packets related to a flow/object and forward them at a later time. Time delay to transmission from storage may be short (Micro Time Shift) or long (Macro Time Shift). Their slides are here. This was a joint presentation by Tom & Shyam from AT&T, Bapi Vinnakota from Netronome. These are early ideas. Tom, Shyam and Bapi offered to do a second presentation on this, diving deeper into data plane constructs if any new support is needed from P4 architecture/language.
Mickey Spiegel from Barefoot Networks presented initial thoughts on an Out-of-Band mechanism for communicating metadata semantics between INT switches and INT monitoring system. The mechanism is based on addition of an Extensions Header in Telemetry Reports. Mickey's presentation can be accessed here.
This is an excellent start and the high-level approach made sense to all attendees. Mickey also did a great job of capturing all aspects of metadata semantics that potentially need to be reported (See slide 4 in his slide deck). Some questions that came up with respect to this proposal -
- Should metadata semantics be reported over UDP or TCP. Can we simply define the packet formats and leave it up to individual switch and monitoring system implementations to decide what transport protocol(s) they support.
- Should metadata semantics be on a per-switch basis or a per-line-card basis. Telemetry report header supports per-line-card reporting by means of including Switch ID and a "HW ID" in the header. However, INT data plane hop-by-hop header only includes Switch ID. So there is some disconnect here, which may be addressed by embedding "HW ID" in level 1/level 2 port ID in INT data plane.
- How to structure the metadata semantics in the Telemetry Report Extensions Header - TLV/JSON/Protobuf/Yang?
- Next working group meeting will be on May 31, 2018. Agenda and location to be announced closer to the date.
- It is important to close on the structure of the Telemetry Report Extensions Header, with respect to whether we want to use TLVs/JSON/Protobufs/Yang to make further progress on metadata semantics reporting. Let us try and exchange thoughts over this on the mailing list to speed things up and not wait until the next working group meeting. If we can close on the language for schema modeling, it will help Mickey on defining further details of the Extension Header format.