Dynamic Expert Policy Selection for Thermal Management in MPSoCs
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional dynamic thermal management (DTM) techniques in multiprocessor system-on-chips (MPSoCs) fail to balance temperature across the chip, leading to thermal hot spots, increased load average, and reliability issues, while also exacerbating thermal cycling, which can cause permanent failures.
Innovation Solution
A system that dynamically selects an expert policy based on monitored performance parameters, using an online learning technique and loss function to optimize operational metrics such as load average, reliability, and energy efficiency, by systematically tracking temporal relationships between performance parameters and selecting the most suitable policy to manage thermal hot spots, cycles, spatial gradients, and load average.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Temperature
If conventional dynamic thermal management (DTM) techniques are used to prevent thermal hot spots, then temperature stability is improved, but load average increases
Solution Approach 1:
The patent applies dynamics by transitioning from static threshold-based thermal management to dynamic policy selection. The system dynamically selects from multiple expert policies based on real-time performance parameters, allowing adaptive response to varying thermal and workload conditions. This enables the system to optimize between temperature control and load average differently under various conditions.
Solution Approach 2:
The patent changes parameters by monitoring multiple performance parameters (temperature, load average, thermal cycling) and using them to select different expert policies. The system adjusts its behavior based on the current state of these parameters, switching between policies that prioritize temperature control versus those that prioritize productivity, thereby resolving the contradiction dynamically.
2Reliability
If conventional DTM techniques are used to maintain safe temperatures, then reliability is improved, but thermal cycling increases
Solution Approach 1:
The patent implements feedback by continuously monitoring performance parameters including thermal cycling metrics and using this information to select expert policies. The feedback loop allows the system to detect when thermal cycling is occurring and adjust policy selection to reduce harmful thermal cycles while maintaining reliability through temperature control.
Solution Approach 2:
The system dynamically adapts its thermal management strategy based on real-time conditions. By selecting from multiple expert policies based on current performance parameters, the system can reduce thermal cycling during certain conditions while maintaining temperature control during others, thereby reducing harmful factors without sacrificing reliability.
3Temperature
If expert policies optimize for temperature control, then temperature profile is improved, but adaptability to varying workloads decreases
Solution Approach 1:
The patent applies dynamics through dynamic policy selection based on workload characteristics. The system monitors performance parameters that reflect workload conditions and selects expert policies accordingly. This enables the system to adapt to varying workloads by switching between temperature-optimized policies and productivity-optimized policies, thereby maintaining both temperature profile quality and adaptability.
Solution Approach 2:
The system changes its operational parameters by selecting different expert policies based on monitored performance parameters. This parameter change approach allows the system to adapt to varying workload conditions while maintaining good temperature profiles, as the policy selection is driven by real-time system state information.
Data Source
AI summary
Some embodiments of the present invention provide a system that manages a performance of a computer system. During operation, a current expert policy in a set of expert policies is executed, wherein the expert policy manages one or more aspects of the performance of the computer system. Next, a set of performance parameters of the computer system is monitored during execution of the current expert policy. Then, a next expert policy in the set of expert policies is dynamically selected to manage the performance of the computer system, wherein the next expert policy is selected based on the monitored set of performance parameters to improve an operational metric of the computer system.