Computational Storage Data Transfer via Peer-to-Peer Offload
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Computational storage devices face limitations in scaling due to PCIe bandwidth constraints and memory address space limitations, particularly when multiple devices are used in servers, leading to restricted memory mapping and potential performance bottlenecks.
Innovation Solution
Implementing a mechanism for efficient device memory management that allows data processing functions to be offloaded to storage devices, enabling peer-to-peer data transfers without relying on host system address space mapping, using techniques like scatter gather lists and device-managed memory allocation to facilitate internal data transfers and scaling.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If PCIe bandwidth is used for data transfer between host and storage devices, then data transfer capability is provided, but bandwidth constraints limit scaling to multiple devices
Solution Approach 1:
The patent introduces peer-to-peer data transfer as an intermediary mechanism that enables direct communication between storage devices and host memory without relying on PCIe bandwidth. This allows multiple storage devices to scale beyond PCIe constraints by establishing direct memory access paths.
Solution Approach 2:
The patent transitions from traditional PCIe-based data transfer to peer-to-peer memory access, representing a dimensional shift in the data transfer architecture. This enables scaling to multiple devices by creating parallel data transfer paths that bypass the PCIe bandwidth bottleneck.
2Ease of operation
If host system address space mapping is used for memory access, then data transfer is enabled, but memory address space limitations restrict the number of supported devices
Solution Approach 1:
The patent segments the address space management by allowing each storage device to maintain its own independent address space. This eliminates the host system address space limitation by dividing the addressing functionality into device-specific segments that can be accessed directly through peer-to-peer transfers.
Solution Approach 2:
The patent uses device-managed memory allocation as an intermediary mechanism that enables data transfer without requiring host system address space mapping. Each device manages its own memory allocation and addressing, allowing multiple devices to operate independently beyond host address space constraints.
3Productivity
If multiple storage devices are deployed to improve processing capacity, then productivity increases, but PCIe bandwidth constraints create performance bottlenecks
Solution Approach 1:
The patent introduces peer-to-peer memory access as a new dimension for data transfer that bypasses the PCIe bandwidth constraint. This allows multiple storage devices to be deployed for increased processing capacity without being limited by PCIe bandwidth availability.
Solution Approach 2:
The patent establishes peer-to-peer data transfer as an intermediary path that enables direct memory access between storage devices and host, eliminating the PCIe bandwidth bottleneck that would otherwise limit performance when multiple devices are deployed.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
Provided are systems, methods, and apparatuses for managing memory. The method can include: establishing a connection via an interface, between a host device and a storage device; and transferring data, via the interface, between first memory associated with the host device and second memory associated with the storage device by performing a data operation on the second memory by an application executed by the host, where the storage device includes a processing element that accelerates the data operation by performing at least one offload function on the data operation.