RDMA Latency Reduction via Intermediate Buffer Overlap
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The existing Remote Direct Memory Access (RDMA) protocol introduces latency due to the registration and deregistration of application memory buffers, which slows down information transfer between networked devices, as the operating system kernel may move data to virtual memory, causing corruption and requiring buffer registration and deregistration to prevent this.
Innovation Solution
The proposed solution involves transferring information to intermediate buffers during registration and deregistration, allowing the transfer process to overlap with these tasks, reducing the amount of data to be transferred while registered and utilizing time that would otherwise be idle, thus minimizing latency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If the operating system kernel moves data to virtual memory to free up physical memory, then physical memory resources are freed for other uses, but the application memory buffer must be registered and deregistered which introduces latency into RDMA
Solution Approach 1:
The system performs preliminary registration of application memory buffers with the RDMA subsystem before actual data transfer occurs. This advance registration ensures the buffers are locked and protected from being moved to virtual memory, eliminating the need for registration/deregistration during active transfer and reducing RDMA latency.
Solution Approach 2:
The system maintains continuous RDMA operations by keeping memory buffers registered and locked throughout the entire data transfer process. Instead of registering and deregistering buffers intermittently, the buffers remain in a stable registered state, allowing uninterrupted direct memory access operations without latency interruptions.
2Reliability
If memory buffers are registered to prevent contents from being moved during RDMA transaction, then data integrity is maintained, but the registration/deregistration process introduces latency that slows down information transfer
Solution Approach 1:
Memory buffers are registered with the RDMA subsystem in advance before data transfer begins. This preliminary registration locks the buffers and prevents the operating system kernel from moving their contents to virtual memory, ensuring data integrity is maintained throughout the transfer without requiring repeated registration/deregistration cycles that would introduce latency.
Solution Approach 2:
Buffers remain in a continuously registered and locked state throughout the entire RDMA transfer operation, maintaining data integrity without interruption. The continuous registered state eliminates the need for repeated registration/deregistration operations, thereby maintaining reliability while minimizing latency.
Data Source
AI summary
The present disclosure provides systems and methods for remote direct memory access (RDMA) with reduced latency. RDMA allows information to be transferred directly between memory buffers in networked devices without the need for substantial processing. While RDMA requires registration/deregistration for buffers that are not already preregistered, RDMA with reduced latency transfers information to intermediate buffers during registration/deregistration, utilizing time that would have ordinarily been wasted waiting for these processes to complete, and reducing the amount of information to transfer while the source buffer is registered. In this way the RDMA transaction may be completed more quickly. RDMA with reduced latency may be employed to expedite various information transactions. For example, RMDA with reduced latency may be utilized to stream information within a device, or may be used to transfer information for an information source external to the device directly to an application buffer.