Multipath Transport for RDMA Networking

PhD Qualifying Examination


Title: "Multipath Transport for RDMA Networking"

by

Mr. Xiangzhou LIU


Abstract:

Remote Direct Memory Access (RDMA) is increasingly used in production 
datacenters for its high throughput, low latency, and low CPU overhead. 
Datacenter fabrics provide multiple paths, but conventional RDMA transport 
typically assigns each flow to a single path. This can cause congestion on 
heavily loaded paths while leaving others underutilized. Multipath transport 
can improve fabric utilization by distributing traffic across these paths and 
adjusting path usage in response to congestion. Realizing this benefit in RDMA 
requires load balancing mechanisms with low memory and processing overhead on 
RDMA network interface cards (RNICs) and network switches.

This survey reviews in-network and host-side load balancing mechanisms for 
multipath RDMA, comparing their effectiveness in utilizing network capacity 
and adapting to congestion. The comparison accounts for the memory and 
processing overhead on RNICs and network switches, highlighting the tradeoffs 
between load balancing performance and hardware cost. Advances in multipath 
load balancing call for corresponding adaptations to congestion control and 
reliability mechanisms. The survey therefore examines how congestion control 
coordinates sending rates with traffic distribution and how reliability 
mechanisms tolerate packet reordering while recovering losses promptly.

The survey identifies common challenges in sustaining effective multipath 
transport as the number of active connections and network size grow. It 
outlines research opportunities in hardware-efficient designs for load 
balancing, congestion control, and reliability in next-generation RDMA 
transport.


Date:                   Monday, 26 October 2026

Time:                   9:00am - 11:00am

Venue:                  Room 5504
                        Lifts 25/26

Committee Members:      Prof. Kai Chen (Supervisor)
                        Prof. Song Guo (Chairperson)
                        Dr. Xiaomin Ouyang