Data Throughput Estimation Model for Computing Systems
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for predicting application performance in computing systems from resource statistics are inefficient and lack accuracy, making it difficult to detect bottlenecks and optimize resource utilization.
Innovation Solution
A data throughput estimation model is generated using machine learning algorithms, such as decision trees, boosting, and random forests, which correlates data throughput with various resource statistics like thread migrations, inter-processor cross-calls, and system calls, enabling accurate prediction of throughput.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional monitoring methods are used to measure resource statistics, then implementation is simple, but prediction accuracy of application performance is poor
Solution Approach 1:
The patent transforms the prediction approach by changing from direct traditional monitoring to a machine learning-based parameter transformation. It uses multiple resource statistics parameters (CPU utilization, memory usage, disk I/O, network throughput) as inputs to a trained model that outputs predicted application performance metrics, thereby achieving higher prediction accuracy through parameter transformation rather than direct measurement.
Solution Approach 2:
The patent applies preliminary action by pre-training the machine learning model offline using historical data before actual prediction. The model is trained in advance with labeled historical resource statistics and corresponding application performance data, so that during runtime, only the trained model needs to be applied without requiring real-time complex computations, thus resolving the contradiction between accuracy and complexity.
2Measurement precision
If machine learning models are trained offline and applied online, then prediction accuracy improves, but training time and data requirements increase
Solution Approach 1:
The patent separates the prediction process into offline training phase and online application phase. During offline training, the machine learning model learns from historical data. During online operation, only the pre-trained model is applied to new data without requiring retraining, thus achieving high prediction accuracy while minimizing real-time computational overhead and training time.
Solution Approach 2:
The patent segments the overall prediction task into two distinct phases: an offline training phase where the model is built using historical data, and an online prediction phase where the trained model is applied. This segmentation allows computationally intensive training to occur when time is not constrained, while online prediction operates efficiently with minimal time loss.
Data Source
AI summary
Embodiments include systems and methods for generating a data throughput estimation model. A system may be monitored to measure both (a) data throughput and (b) computing statistics of one or more computing resources to generate an initial data set. The relationship between the data throughput and the computing statistics, in the initial data set, is used to generate a data throughput estimation model. The data throughput estimation model may be generated using a machine learning model, a neural network algorithm, boosting decision tree algorithm, and/or a random forest decision tree algorithm. Additional measurements of the computing resource statistics may be applied to the data throughput estimation model to estimate data throughput.