Storage Datapath ML Inference Using Sampled IO Feature Matrices
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The cost of generating features and building machine learning models in storage systems with millions of objects is prohibitive due to memory footprint and CPU overhead, impacting I/O processing performance.
Innovation Solution
A method that samples IO requests from processing cores, generates an IO feature matrix, and processes it using a machine learning model separate from the storage system datapath, providing the result to execute storage system policies.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If machine learning models are trained using real-time IO request sampling in storage systems with millions of objects, then model accuracy and storage optimizations are improved, but memory footprint and CPU overhead become prohibitive
Solution Approach 1:
The patent segments the storage system into two independent parts: the datapath that processes IO requests and the analytics engine that trains machine learning models. This segmentation allows the analytics engine to access IO request data without being integrated into the critical IO processing path, thereby reducing memory footprint and CPU overhead while maintaining model accuracy.
Solution Approach 2:
The patent implements preliminary action by having the analytics engine proactively pull IO request data from the datapath at predetermined intervals and pre-process it into feature matrices before model training. This approach allows the system to prepare training data in advance without impacting real-time IO performance, resolving the contradiction between accurate modeling and resource consumption.
2Productivity
If machine learning models are trained using real-time IO request sampling, then storage optimizations are improved, but I/O processing performance (throughput and latency) deteriorates
Solution Approach 1:
The patent introduces an intermediary mechanism where the analytics engine independently pulls IO request data from the datapath at predetermined intervals. This intermediary approach decouples the model training process from the IO processing path, allowing storage optimizations to be improved through accurate modeling without deteriorating I/O processing speed.
Solution Approach 2:
The patent implements periodic action by having the analytics engine sample and process IO requests at predetermined intervals rather than in real-time. This periodic processing allows the system to maintain high I/O processing speed while still gathering sufficient data for effective machine learning model training and storage optimizations.
3Measurement precision
If feature matrices are generated and processed for machine learning models, then model inference accuracy is improved, but computation overhead increases
Solution Approach 1:
The patent applies preliminary action by pre-processing IO request data into feature matrices in advance, before model inference is needed. This pre-processing approach allows the system to perform computationally intensive feature extraction and matrix generation beforehand, reducing computation overhead during actual model inference while maintaining high inference accuracy.
Solution Approach 2:
The patent segments the computation into distinct phases: data collection, feature extraction, matrix generation, and model inference. By separating these phases and performing them at different times and on different systems (datapath vs. analytics engine), the patent reduces computation overhead during critical operations while maintaining model inference accuracy.
4Measurement precision
If IO request data is collected and processed in real-time, then model training quality is improved, but system complexity increases
Solution Approach 1:
The patent segments the system into independent components with clear responsibilities: the datapath handles IO request processing while the analytics engine handles data collection and model training. This segmentation improves model training quality by providing dedicated resources for analytics while reducing system complexity through clear separation of concerns and independent operation of each component.
Data Source
AI summary
A method, computer program product, and computing system for processing a plurality of input/output (IO) requests on a storage object within a storage system datapath using a plurality of processing cores of a storage system. The plurality of IO requests are sampled from the plurality of processing cores. An IO feature matrix is generated using the periodically sampled IO requests. The IO feature matrix is provided to a machine learning model that is separate from the storage system datapath. A machine learning model inference result is generated by processing the IO feature matrix using the machine learning model. The machine learning model inference result is provided to the storage system datapath. A storage system policy is executed using the machine learning model inference result.