Multicore Performance Sampling With Lower Bus Bandwidth
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current multicore processor systems face inefficiencies in data sampling and transfer due to high sampling rates required by the Shannon/Nyquist theorem, leading to excessive processor-memory bus usage and heat generation, especially when dealing with large numbers of cores, which complicates task scheduling and performance monitoring.
Innovation Solution
Implementing compressive sampling techniques that allow for the direct compression of processor performance data during the sampling phase, reducing the number of measurements needed to reconstruct the original signal, thereby minimizing bus bandwidth usage and heat generation, and enabling efficient data transfer and scheduling.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional uniform sampling is used to monitor processor performance, then complete information is captured, but the data bus bandwidth and processing cycles are excessively consumed
Solution Approach 1:
The patent extracts only the essential performance information from the processor cores using compressive sampling, rather than transferring all sampled data. The PMU captures performance metrics and applies compressive sampling to extract the most significant information, reducing the data volume transferred over the data bus while maintaining the ability to reconstruct accurate performance profiles.
Solution Approach 2:
The patent changes the sampling parameters by using compressive sampling techniques that allow reconstruction of the full performance signal from a reduced number of measurements. By transforming the sampling approach from uniform high-rate sampling to compressive sampling with fewer measurements, the system maintains measurement precision while reducing data quantity.
2Measurement precision
If the sampling rate is increased to capture complete processor performance data, then measurement accuracy is improved, but the complexity and cost of the sampling system increases
Solution Approach 1:
The patent introduces compressive sampling as an intermediary process between the Performance Monitor Unit and the scheduler. The PMU captures performance data and applies compressive sampling algorithms to reduce the data dimensionality before transfer. This intermediary step allows the system to maintain high measurement precision while using a simpler, lower-bandwidth data transfer mechanism.
3Loss of information
If traditional two-step sampling and compression is used, then data is captured completely, but processing time and bus cycles are wasted
Solution Approach 1:
The patent merges the sampling and compression steps into a single compressive sampling operation. Instead of first capturing complete performance data and then compressing it, the system performs compressive sampling that directly produces a reduced representation of the performance data. This unified approach eliminates the separate compression step and reduces the time spent on data processing and transfer.
Data Source
AI summary
Methods and apparatus for gathering information from processors by using compressive sampling are presented. The invention can monitor multicore processor performance and schedule processor tasks to optimize processor performance. Using compressive sampling minimizes processor-memory bus usage by the performance monitoring function. An embodiment of the invention is a method of gathering information from a processor, the method comprising compressive sampling of information from at least one processor core. The compressive sampling produces compressed information. The processor comprises the at least one processor core, and the at least one processor core is operative to process data.