Memory Usage Determination via Robust Regression
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Cloud service providers face challenges in predicting and detecting Service Level Agreement (SLA) violations due to inadequate tools, which can lead to delayed resolution and adverse customer experiences, and require burdensome data analysis to ensure compliance.
Innovation Solution
The use of robust least-squares regression techniques to estimate seasonal indices for multiple periods, allowing for the identification of code blocks for optimization and forecasting of heap usage and thread intensity without instrumentation, thereby providing proactive monitoring and reducing overhead.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If instrumentation code is added to collect measurements of low-level events, then measurement precision is improved, but device complexity and execution time increase
Solution Approach 1:
The patent introduces an intermediary sampling mechanism that collects measurements without requiring instrumentation code to be embedded in the monitored system. The sampler acts as a separate entity that periodically captures state information (stack traces, heap usage) from the system, thereby achieving measurement precision without adding complexity to the original system architecture
Solution Approach 2:
The patent replaces the mechanical approach of inserting instrumentation code directly into the system with a statistical sampling approach. Instead of instrumenting every event, the system uses periodic sampling of system state combined with probabilistic data structures (Count-Min Sketch) to estimate measurements, substituting direct mechanical measurement with statistical inference
2Measurement precision
If instrumentation code is added to collect measurements, then measurement precision is improved, but productivity decreases due to increased execution time
Solution Approach 1:
The patent applies partial action by sampling only a subset of system states rather than measuring every event. The sampling rate is tuned to provide sufficient measurement precision for detecting SLA violations while minimizing the overhead impact on system productivity. This selective measurement approach ensures that measurements are accurate enough for their purpose without unnecessarily burdening the system
Solution Approach 2:
The patent replaces the time-consuming instrumentation approach with efficient statistical sampling and probabilistic data structures. The Count-Min Sketch algorithm provides space-efficient and time-efficient estimation of heap usage and other metrics, achieving measurement precision without the execution time penalty of traditional instrumentation
3Ease of operation
If traditional regression techniques are used to forecast memory usage, then ease of operation is maintained, but reliability decreases due to weak-exogeneity and heteroscedasticity
Solution Approach 1:
The patent transforms the regression problem by changing the parameterization approach. Instead of using traditional OLS regression that assumes constant variance and exogeneity, the patent applies robust regression techniques (Huber loss, LAD) that are resistant to heteroscedasticity and weak-exogeneity. This parameter change in the loss function allows the model to handle the actual statistical properties of memory usage data while maintaining operational simplicity through automated model selection
Data Source
Figure 1
Figure 2
Figure 3
AI summary
Embodiments provide techniques for estimating seasonal indices for multiple periods. Some embodiments can receive a signal comprising a plurality of measures sampled over a span of time from an environment in which one or more processes are being executed. Some embodiments may then extract a seasonal effector and a de-seasonalized component from the signal. Next, some embodiments can apply one or more spline functions to the seasonal effector to generate a first model. Some embodiments may then apply a linear regression technique to the de-seasonalized component to generate a second model. Some embodiments may then initiate actions associated with the code. Some embodiments may then generate a forecast of the signal based on the first model and the second model. Next, some embodiments may initiate, based at least in part on the forecast, one or more actions associated with the environment.