Task Instance Outlier Detection in Parallel Processing Systems
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In large-scale parallel processing systems, task instances often interfere with each other due to unequal resource usage, leading to increased latency and inefficiency, as they share common resources like CPU and memory.
Innovation Solution
A method is implemented to identify poorly performing task instances by analyzing metric data such as cycle per instruction (CPI) and cache usage, determining deviations from the mean, and generating a report to rank and potentially modify or relocate these instances to improve overall system performance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Use of energy by moving object
If task instances share common resources (CPU, memory) in a parallel processing system, then system resource utilization is improved, but task instance performance deteriorates due to interference and increased latency
Solution Approach 1:
The system continuously collects metric data (CPI, cache usage) from task instances, compares each instance's performance against the mean and standard deviation of all instances, and provides feedback by identifying outliers. This feedback mechanism enables the system to detect and address performance degradation caused by resource sharing interference.
Solution Approach 2:
The patent extracts and isolates poorly performing task instances (outliers) from the parallel processing system by identifying them through statistical deviation analysis. These extracted outliers can then be relocated or terminated, removing the source of interference from the shared resource environment.
2Productivity
If more task instances are executed in parallel on a cluster, then system productivity is improved, but the number of interfering task instances increases leading to higher latency
Solution Approach 1:
The system implements continuous monitoring of task instance performance metrics and compares them against statistical thresholds. This feedback loop enables real-time detection of latency increases and performance degradation, allowing the system to maintain high productivity while managing the time loss introduced by parallel execution interference.
Solution Approach 2:
By extracting and isolating outlier task instances that cause excessive interference, the system preserves the high productivity benefits of parallel execution while removing the sources of latency. This allows more task instances to run in parallel without proportionally increasing overall system latency.
3Power
If task instances with higher resource requirements are allowed to execute, then computational capability is improved, but other task instances experience increased latency due to resource contention
Solution Approach 1:
The system monitors resource usage patterns and performance metrics of all task instances, providing feedback that identifies which instances are consuming excessive resources. This enables the system to maintain high computational capability while detecting and addressing the latency impact on other instances through statistical outlier detection.
Solution Approach 2:
The patent extracts and isolates task instances that consume disproportionate resource shares, removing them from the parallel execution environment. This extraction preserves the overall computational capability of the system while eliminating the resource contention that causes increased latency for other legitimate task instances.
Data Source
AI summary
Among other disclosed subject matter, a method includes receiving metric data associated with an execution of each of a plurality of task instances. The plurality of task instances include task instances associated with a task and the metric data for each task instance relating to execution performance of the task instance. The method includes for each task instance determining a deviation of the metric data associated with the task instance relative to an overall deviation of the metric data for the plurality of task instances of the task during each of a plurality of intervals and combining deviation measurements for the task instance that exceed a threshold deviation to obtain a combined deviation value. Each deviation measurement corresponds to the deviation of the metric data for one of the plurality of intervals. The method includes ranking the combined deviation values associated with at least a subset of the task instances.