Queuing Model Estimation via Response Time Analysis
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for estimating service demands in software and hardware systems, such as CPU utilization, are inaccurate and difficult to implement, especially in virtualized environments and when data is not publicly available, leading to inadequate performance modeling.
Innovation Solution
A computer-implemented method that generates a queuing model by measuring response times and estimating mean service demands using linear regression and maximum likelihood analyses, allowing for the parameterization of queuing models with characteristics like scheduling and service demand distributions, which can be applied to any computer system, including those owned by third parties.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If CPU utilization measurement is used to estimate service demands, then service demand estimation can be obtained, but the measurement accuracy deteriorates due to operating system access requirements, specialized instrumentation interference, and inability to access data in virtualized environments or third-party systems
Solution Approach 1:
The patent introduces response time as an intermediary measurement that indirectly reflects service demands without requiring direct CPU utilization access. By measuring the total time from request submission to completion, the system can estimate service demands through statistical analysis of response time distributions, bypassing the need for operating system access or specialized CPU instrumentation.
Solution Approach 2:
The patent replaces the mechanical CPU sampling instrumentation approach with a software-based response time measurement system. Instead of using hardware-level CPU counters and samplers that require OS access, the system uses application-level timing mechanisms to measure response times, which can be implemented in any system regardless of virtualization or ownership constraints.
2Reliability
If CPU sampling instrumentation is used to measure utilization, then service demand data can be collected, but system reliability deteriorates due to interference with normal system activities and inaccurate measurements in virtualized environments
Solution Approach 1:
The patent implements a self-service measurement approach where the system being measured provides its own response time data without external instrumentation. The application or service itself can log response times or make them available through standard interfaces, eliminating the need for complex external CPU sampling instrumentation that interferes with system operations and provides unreliable data in virtualized environments.
3Adaptability or versatility
If traditional CPU utilization methods are used, then service demand estimation is possible, but adaptability deteriorates because the method cannot be applied to virtualized environments, third-party systems, or systems without OS access
Solution Approach 1:
The patent creates a universal response time measurement approach that can be applied across diverse system types including physical servers, virtualized environments, cloud services, and third-party systems. By measuring response times at the application level rather than CPU level, the same methodology works universally regardless of underlying infrastructure, while maintaining estimation accuracy through statistical analysis of response time distributions.
Data Source
AI summary
Implementations of the present disclosure provide computer-implemented methods including defining a workload comprising a plurality of service requests, each service request corresponding to a class of a plurality of classes, applying the workload to a computer system that receives and processes service requests, measuring a response time of the computer system for each request of the workload, estimating a mean service demand for each class based on the response times and a base queuing model that represents the computer system, and generating the queuing model based on the mean service demands and characteristics of the workload.