TL;DR
This paper introduces MonTM, a hardware-based thermal management system for mixed-criticality systems (MCSs) on many-core processors. MonTM prevents safety-critical tasks from missing deadlines due to thermal interference from best-effort tasks by using a novel interconnect for thermal pre-error communication and a slack monitor to dynamically adjust task frequencies, leading to significant performance improvements for best-effort tasks.
What Problem Does This Paper Address?
Modern embedded systems are increasingly integrating tasks with varying safety requirements (e.g., safety-critical functions in autonomous driving alongside less critical infotainment systems) onto single many-core processors. This creates "mixed-criticality systems" (MCSs). While isolating shared architectural resources (like caches or memory) is a known challenge, this paper focuses on a less obvious but equally critical interference channel: the processor's thermal behavior. When high-performance best-effort tasks run on cores adjacent to safety-critical tasks, their heat generation can cause thermal violations, leading to throttling that can make safety-critical tasks miss their deadlines. Existing thermal management techniques often err on the side of caution, potentially underutilizing resources. This paper tackles the problem of efficiently managing thermal resources in MCSs to guarantee safety-critical task deadlines while maximizing the performance of best-effort tasks.
Methods
The core of MonTM is a two-pronged approach:
- Hardware-Based Thermal Interconnect: MonTM introduces a lightweight, hardware-based interconnect that allows thermal managers (DTMs) on different cores to communicate "thermal pre-error" signals. These signals indicate an imminent thermal violation on a safety-critical core. This allows neighboring best-effort tasks to proactively throttle their performance before a critical deadline is missed. The urgency of the pre-error is communicated through different levels, dictating how aggressively best-effort tasks should be throttled.
- Slack Monitoring for Safety-Critical Tasks: A novel "slack monitor" is implemented in hardware. This monitor tracks the progress of safety-critical tasks and their remaining execution time relative to their deadlines. If a safety-critical task is progressing faster than its worst-case execution time (WCET) predicts (i.e., it has "slack"), its voltage/frequency (V/f) level can be safely reduced. This reduction frees up thermal headroom, which can then be utilized by best-effort tasks, boosting their performance without compromising safety.
Datasets
The paper does not explicitly mention using pre-existing datasets. Instead, they generate synthetic workloads. These workloads are created by embedding code blocks with varying run-time properties into random control flow graphs (CFGs). To validate their approach, they also use a real-world application benchmark: an object detection chain consisting of multiple computer vision algorithms. This synthetic workload generation allows them to create diverse load scenarios and power consumption variances, mimicking real-world complexities.
Key Insights and Ideas
- Thermal Interference is a Major Bottleneck in MCSs: The paper highlights that thermal coupling between cores is a significant challenge in MCSs, often overlooked compared to resource contention.
- Proactive Thermal Management is Crucial: Relying solely on reactive thermal throttling (i.e., reacting only when a violation occurs) is insufficient for MCSs. Proactive communication of imminent thermal issues is key.
- Dynamic V/f Scaling Based on Slack is Highly Effective: Static V/f assignments for safety-critical tasks based on WCET can be overly conservative. Dynamically adjusting these levels based on actual task progress (slack) unlocks significant thermal headroom.
- Hardware Implementation is Key for Performance: The proposed solutions (interconnect and slack monitor) are implemented in hardware. This is crucial for achieving the low latency and high throughput required for effective real-time thermal management.
Key Results
- MonTM successfully guarantees that safety-critical tasks meet their deadlines while simultaneously improving the performance of best-effort tasks.
- Compared to state-of-the-art techniques like GDP and PdRM, MonTM reduces the average run-time of best-effort tasks by 7%-44% without slack monitoring.
- With the addition of the slack monitor, this reduction is further improved by an additional 1%-6%, with some individual best-effort tasks seeing reductions of up to 12%.
- In their FPGA prototype evaluation, MonTM achieved up to a 45% reduction in best-effort task run-time compared to the state of the art.
- The hardware overhead of MonTM is minimal, with a total of 1.3% of slice LUTs and 4.4% of slice registers on their FPGA prototype, which is considered well-justified by the performance gains.
Limitations and Future Work
- Pessimism in Safety-Critical Task V/f: While MonTM reduces V/f for safety-critical tasks with slack, the paper notes that in scenarios with no best-effort tasks nearby, a "race to idle" strategy with the original frequency might be more beneficial to minimize future overlaps. The current implementation considers this trade-off by extending the DTM interconnect to communicate best-effort task activity.
- Synthetic Workloads: While synthetic workloads are useful for controlled evaluation, their direct correlation to real-world application power consumption is an ongoing area of research. The paper acknowledges that synthetic benchmarks serve as a conservative replication.
- Future Work: The authors suggest further exploring the trade-off between proactive thermal management and the "race to idle" strategy. Additionally, extending the approach to more complex many-core architectures and heterogeneous systems could be valuable. The paper also hints at the potential for further optimization of the slack monitoring mechanism.
Summary updated