Storage Controller Asynchronous Event Reporting for Kernel Status Monitoring
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for monitoring the status of acceleration kernels in storage devices lead to PCIe bus saturation and increased CPU overhead, causing inefficiencies and potential bottlenecks in system performance due to direct communication between kernels and the host.
Innovation Solution
A method where the storage device's controller receives asynchronous event reporting commands, builds a look-up table, and sends command completion entries to the host only when matching completion messages are received, thereby preventing direct kernel-to-host communication and reducing PCIe bus load, while also monitoring temperature to delay entries if necessary.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If acceleration kernels directly communicate with the host to report status, then the host can monitor kernel status in real-time, but the PCIe bus becomes saturated and CPU overhead increases
Solution Approach 1:
The patent introduces an intermediary buffer memory and controller between the acceleration kernels and the host. The buffer stores completion messages from kernels, and the controller manages the asynchronous event reporting process, preventing direct kernel-to-host communication that would saturate the PCIe bus while still enabling accurate status monitoring.
Solution Approach 2:
The system pre-allocates buffer memory and sets up event reporting structures before kernels begin execution. Completion messages are pre-stored in the buffer as they are generated by kernels, allowing the host to retrieve status information without causing bus saturation during active computation.
2Measurement precision
If the host polls acceleration kernels frequently to check status, then real-time monitoring is achieved, but PCIe bandwidth is wasted and CPU overhead increases
Solution Approach 1:
The system implements asynchronous event reporting where completion messages are automatically fed back to the host through the buffer and controller when ready, eliminating the need for continuous host polling. This feedback mechanism provides timely status updates while minimizing PCIe bandwidth consumption.
Solution Approach 2:
Instead of continuous polling, the host receives periodic status updates only when completion messages are available in the buffer. This transforms continuous polling into event-driven periodic communication, reducing unnecessary PCIe transactions while maintaining monitoring effectiveness.
3Speed
If completion messages are sent immediately when kernels finish operations, then real-time status reporting is achieved, but PCIe bus saturation occurs
Solution Approach 1:
Multiple completion messages from different kernels are merged and batched together in the buffer before being transmitted to the host. This combining of individual status reports into consolidated updates maintains fast status reporting while significantly reducing the total number of PCIe transactions and preventing bus saturation.
Solution Approach 2:
Completion messages are preliminarily stored in the buffer as soon as kernels complete their operations, rather than being immediately transmitted. This preliminary storage allows for message coalescing and timing optimization, enabling fast status capture while deferring transmission to optimize PCIe bus usage.
Data Source
AI summary
A method for status monitoring of acceleration kernels in a storage device is provided. The method includes: receiving an asynchronous event reporting (AER) command by a controller of the storage device from a host, the AER command corresponding to a first operation assigned to a first acceleration kernel on an acceleration co-processor by the host; adding the received AER command to a look-up table in the controller; receiving a completion message from the first acceleration kernel corresponding to the first operation; comparing the received completion message to the AER commands in the look-up table; and when a match is found between the received completion message and one of the AER commands in the look-up table, sending a command completion entry to the host.