Multi-Path Traffic Engineering Workshop#

Vishnu Pavan Kumar Beeram – HPE, Zhaohui (Jeffrey) Zhang – HPE, Kireeti Kompella – Juniper Networks

• Why: MPTE unifies multipath and TE for WAN and AIDC/DCI, and is feeding into ML network scheduling at the IETF — relevant to the PM's WAN/interconnect and AI-fabric interests.
• What: One MPTE DAG construct carries what previously needed many LSPs: unequal-cost, weighted load balancing at every junction, multiple ingresses and egresses, minimal churn on link failure (junction ratios adjust, DAG shape stays), and much less state than per-path signaled tunnels. For ML clusters: non-multipath TE is called "a total non-starter," and MPTE becomes the vehicle for network scheduling — resource reservation, protection, and traffic isolation per AI job (draft-kompella-rtgwg-mlnwsched, new at RTGWG).
• How: MPTED tunnel/junction state (incoming bandwidth, previous/next-hop sets with load shares), signaled via RSVP or BGP extensions; the workshop ran it hands-on in HPE/Juniper virtual environments (vMX/cRPD) for both WAN and AIDC scenarios.
• Where applicable: Vendor workshop (HPE/Juniper, registration required) around IETF drafts (draft-kompella-teas-mpte and signaling companions) — standards-track and worth tracking even if you don't run the vendors' stacks; the AIDC framing presumes job-aware scheduling above the fabric.