RDMA Multicast for Reliable Distributed Collective Communication
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current RDMA-based distributed collective communication in AI and HPC applications suffers from high system overheads and bandwidth waste due to point-to-point unicast communication, leading to inefficiencies and increased latency in large-scale computing tasks.
Innovation Solution
Implement multicast solutions using RDMA with a first UD type for multicast packets and a second UD type for unicast acknowledgments, leveraging network switches to reduce redundant data transmission and enhance reliability through real-time data synchronization and feedback mechanisms.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If point-to-point unicast communication is used for distributed collective communication, then communication can be implemented based on existing transport layer connections, but system overheads increase and bandwidth resources are wasted
Solution Approach 1:
The patent introduces a network switch as an intermediary device that enables multicast functionality. The network switch receives unicast packets from senders, identifies multicast groups based on destination addresses, and distributes packets to multiple receivers within the same multicast group. This intermediary approach allows the system to maintain compatibility with existing unicast transport layer connections while achieving efficient multicast communication at the network layer, thereby reducing bandwidth waste without sacrificing adaptability.
2Ease of operation
If point-to-point unicast communication is used for distributed collective communication, then communication can be established using existing connections, but transmission latency increases due to repeated data sending
Solution Approach 1:
The patent merges multiple unicast communication paths into a single multicast communication path by implementing multicast functionality in network switches. Instead of sending separate data packets from each sender to each receiver, the system combines these paths and allows a single packet to be distributed to multiple receivers simultaneously through the multicast-capable network switch, thereby reducing transmission latency while maintaining ease of connection establishment.
Solution Approach 2:
The patent transitions from one-dimensional point-to-point unicast communication to multi-dimensional multicast communication by enabling network switches to handle multicast groups. This dimensional change allows data to flow from one sender to multiple receivers simultaneously through the network switch's multicast forwarding capability, effectively adding a new dimension to the communication topology and reducing the time required for distributed collective operations.
3Adaptability or versatility
If multicast is implemented without network switch support, then multicast functionality can be achieved through software, but system overheads increase
Solution Approach 1:
The patent uses the network switch as an intermediary that offloads multicast processing from end-host software to network infrastructure. The network switch implements multicast group management, packet distribution, and forwarding decisions in hardware, eliminating the need for complex multicast software implementations at senders and receivers. This approach achieves multicast functionality while minimizing system overhead by leveraging the network switch's native capabilities.
Data Source
Figure 1~2
Figure 3A~4
Figure 5A
AI summary
Multicast methods based on RDMA and corresponding node devices are disclosed, which are suitable for distributed collective communication involving a sender and N receivers. According to an example of the methods, the sender sends a multicast packet based on a QP with a first UD type, where the multicast packet carries a first virtual endpoint ID and a first multicast group ID; and each receiver may send a unicast packet to the sender based on a QP with a second UD type, where the unicast packet carries a second virtual endpoint ID and a second multicast group ID. The first multicast group ID of the sender is the same as second multicast group IDs of the N receivers, and the first virtual endpoint ID of the sender and second virtual endpoint IDs of the N receivers are different from each other. In this way, the multicast packet from the sender may be forwarded to all the receivers by using a multicast function of a network switch, and each receiver may reply to the sender with the unicast packet, so as to improve the reliability of distributed collective communication.