Short-Block Read Accumulation for PCIe-Efficient NVMe Storage
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The mismatch between small 'short block' I/O requests in AI workloads and the standard NVMe block read size leads to inefficient use of PCIe bandwidth and degraded system performance due to numerous single-block random reads and increased command processing overhead.
Innovation Solution
Implementing a storage accelerator that aggregates short block I/O requests into a single command, accumulating data into a standard block size and sending it back to the host using a single TLP, reducing the number of commands and optimizing PCIe efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If standard NVMe block read commands are used for short block I/O requests, then data can be read from storage devices, but PCIe bandwidth utilization deteriorates and command processing overhead increases
Solution Approach 1:
The patent merges multiple short block I/O requests into a single NVMe block read command by accumulating data from multiple logical block addresses into one physical block read operation. This combining approach reduces the number of PCIe transactions and command processing overhead while maintaining the ability to serve individual short block requests efficiently
Solution Approach 2:
The system performs preliminary accumulation of short block requests into a buffer before issuing the actual storage read command. By pre-processing and consolidating multiple small requests into a single block-sized request, the system prepares optimized read operations that maximize PCIe bandwidth utilization before data transfer occurs
2Ease of operation
If multiple single-block random reads are performed to satisfy short block requests, then individual data requests can be fulfilled, but command processing overhead increases and system performance degrades
Solution Approach 1:
Multiple individual short block read commands are merged into a single NVMe block read command by accumulating data from multiple logical block addresses (LBAs) into one physical block read operation. This reduces command processing overhead by handling multiple requests through one unified command pathway rather than processing each request separately
Solution Approach 2:
An intermediary accumulation buffer is introduced between the host interface and the storage media. This buffer acts as a mediator that receives multiple short block requests, consolidates them into a single block-sized read operation, and then distributes the accumulated data back to the appropriate destinations, simplifying the command processing pathway
3Productivity
If small TLPs are used to transmit short block data, then data transfer can occur, but PCIe efficiency is reduced
Solution Approach 1:
Multiple small Transaction Layer Packets (TLPs) carrying short block data are merged into a single larger TLP by accumulating the data into a full block size before transmission. This combining of data packets maximizes PCIe lane utilization and improves transmission efficiency by reducing the overhead associated with multiple small packet transactions
Data Source
Figure 1
Figure 2
Figure 3
AI summary
Short block data accumulation techniques with a storage accelerator device may be used to optimize the handling of read requests for data having a size smaller than a block. In one example, a requester (e.g., host) may aggregate requests to read data having a size smaller than a block and send, to a storage accelerator device, a single command (e.g., an NVMe vendor specific command) with information identifying the multiple read requests. The accelerator may then generate the individual block read requests to the SSDs. The storage accelerator may then accumulate the data returned from the storage devices in response to the multiple read requests into a single regular block and send the accumulated data back to the host.