MTP: Transport for In-Network Computing
summary
MTP is a DPDK prototype transport that natively handles pathlet mutation, intercept, reordering and delaying by treating offloads as message-oriented pathlets and using per-pathlet signals (pathlet rxac/txac) plus a dual-RTO scheme to separate fabric vs pathlet delay. In a testbed with a real INC app and 1% random packet loss and 100 μs average pathlet processing, 128 KB messages see far fewer spurious retransmits (negligible above 400 μs fabric RTO). MTP adds resource-specific congestion feedback (8-bit queue reports + Swift-style control), yields more stable convergence than DCTCP, supports exactly-once via virtual channels, and requires no transport state inside pathlets.
AI-written summary · about
chapters (8)
- 0:00MTP and in-network computing
- 1:40Pathlets and overlay paths
- 3:46Limitations of INC transports
- 5:09MTP design and reliability
- 7:41Handling delays with dual RTO
- 9:47Congestion control for pathlets
- 12:15MTP features and evaluations
- 13:00Summary
related talks
32mMPTCP: Extending kernel functionality with eBPF and Netlink
54mHow to [not] implement a [not so] new net protocol
47mMultipath TCP: Breaking Today's Networks with Tomorrow's Protocols
28mMultipath TCP hands-on on Fedora
24mExtending packetdrill to test MPTCP
13mPTYPE Matching as a Fast Way to Steer Traffic
01h 07mLibtpa Introduction
40mTime-Sensitive Networking (TSN) with Mainline Embedded Linux