US 2026/0303686 A1 · ENSURING MEMORY CONSISTENCY IN REMOTE DIRECT MEMORY ACCESS DATA TRANSFERS

Abstract

A system ensures memory consistency during multi-hop intra-node data transfers in a collective communication environment using GPUs, RDMA network interface cards (R-NICs), and high-bandwidth interconnects. It automatically enforces memory ordering across buses such as PCIe and NVLink while remaining transparent to applications. When a receive request completes, the collective transport logic detects a multi-hop path and repurposes the request into a fence or flush command to ensure RDMA-written data is committed and visible. A subsequent gather operation transfers validated message data from the proxy GPU buffer to the destination GPU over the high-bandwidth link. Completion is reported to the application layer with unchanged semantics. These techniques guarantee data integrity, eliminate explicit synchronization calls, and decouple transmit and receive operations from memory fences, improving reliability and communication performance in multi-hop GPU systems.

Claims20

Dependent claims

A method for data transfer using remote direct memory access (RDMA), comprising:

receiving, via an RDMA network interface card (R-NIC), a receive request comprising message data destined for a target processing device, the message data stored in a proxy buffer on a receiving processing device;
determining that the receive request was transmitted via a multi-hop intra-node transport path;
responsive to determining that the receive request was transmitted via a multi-hop intra-node transport path, repurposing the receive request into a memory barrier request, wherein the memory barrier request is configured to commit message data of the receive request to the proxy buffer before further processing of the message data;
performing a gather operation to transfer the message data from the proxy buffer on the receiving processing device to the target processing device;
and
reporting completion of the receive request to an application layer.

3 more elements

A non-transitory computer readable storage medium storing instructions that when executed by one or more computer processors cause the one or more computer processors to perform steps comprising:

receiving, via an RDMA network interface card (R-NIC), a receive request comprising message data destined for a target processing device, the message data stored in a proxy buffer on a receiving processing device;
determining that the receive request was transmitted via a multi-hop intra-node transport path;
responsive to determining that the receive request was transmitted via a multi-hop intra-node transport path, repurposing the receive request into a memory barrier request, wherein the memory barrier request is configured to commit message data of the receive request to the proxy buffer before further processing of the message data;
performing a gather operation to transfer the message data from the proxy buffer on the receiving processing device to the target processing device;
and
reporting completion of the receive request to an application layer.

3 more elements

Dependencies are inferred from the claim text.

Ensuring memory consistency in remote direct memory access data transfers

Life of this patent

Ensuring memory consistency in remote direct memory access data transfers

Life of this patent

Abstract

Claims20