TL;DR
This paper introduces "Quasi Isolation QoS Setups" (QIQoS) to mitigate timing interference in Multi-Processor Systems-on-Chip (MPSoCs) used in safety-critical applications like avionics. By carefully configuring Quality of Service (QoS) features, specifically in the DDR memory controller of the Xilinx Zynq UltraScale+, they achieve strong isolation for critical tasks, reducing performance variations to under 6% even under heavy contention from non-critical tasks.
What Problem Does This Paper Address?
Modern safety-critical systems increasingly rely on Multi-Processor Systems-on-Chip (MPSoCs) to meet demanding performance requirements. These systems often integrate software from multiple vendors into a single architecture (e.g., Integrated Modular Avionics - IMA). A major challenge in these integrated architectures is timing interference: when different software partitions or applications share hardware resources (like the memory controller), their execution can unpredictably impact each other's timing. This makes it difficult to guarantee the real-time performance of critical tasks, complicating software integration, validation, and certification. Existing solutions often struggle with the complexity and limited documentation of modern MPSoCs, leading to overly conservative timing estimates or late discovery of issues. The authors aim to leverage the built-in Quality of Service (QoS) features of MPSoCs to provide a more robust and predictable way to isolate critical tasks from the interference of other applications.
Methods
The core contribution is the concept of Quasi Isolation QoS Setups (QIQoS). Instead of attempting a full, perfect isolation (which is often impractical), QIQoS aims to provide a strong, albeit not full, isolation by strategically configuring existing QoS mechanisms.
The authors instantiate this concept on the Xilinx Zynq UltraScale+ MPSoC, focusing on the DDR Memory Controller (DDRMC), a known bottleneck. They develop two specific QIQoS setups:
- QIQoS1: Leverages transaction timeout features at the port level. Critical tasks (CT) are assigned a low timeout for read requests (VPR traffic class), ensuring they are quickly prioritized if not served. Non-critical tasks (NCT) use higher priority classes (HPR/LPW) but their access is controlled by port throttling based on CAM availability.
- QIQoS2: Focuses on CAM egress control (starvation control). It combines traffic class control with specific settings for the read CAM (rCAM) partitions to ensure CT requests are served promptly, while NCT requests might experience controlled delays.
Both setups utilize traffic classes and port throttling, and importantly, avoid configurations that could lead to incoherent QoS behavior. They also employ reduced command reordering and private ports for critical tasks.
Datasets
Not discussed in the paper. The authors use custom benchmarks and configurations on the Zynq UltraScale+ platform.
Key Insights and Ideas
- QoS as an Isolation Tool: The paper demonstrates that hardware QoS features, often used for performance optimization, can be effectively repurposed for robust timing isolation in mixed-criticality systems.
- "Quasi" Isolation is Sufficient: Achieving perfect isolation is often infeasible due to hardware complexity. The QIQoS concept shows that a carefully engineered "strong" isolation can significantly reduce timing variability and improve predictability.
- DDRMC as a Critical Bottleneck: The focus on the DDRMC highlights its importance as a shared resource where contention significantly impacts real-time performance.
- Configuration is Key: The effectiveness of QIQoS hinges on meticulously selecting and configuring specific QoS parameters (timeouts, throttling, CAM allocation) to balance isolation and performance.
- Early Timing Budgeting: By providing more stable and representative timing figures, QIQoS enables more reliable timing budget consolidation early in the development cycle, reducing integration risks.
Key Results
- Performance Variation Below 6%: Under aggressive QIQoS setups, the performance variation of critical tasks (CT) across different loads from non-critical tasks (NCT) was consistently below 6 percentage points.
- Effective Isolation: The proposed QIQoS configurations successfully isolated the performance of CTs running on the CPU cores (A53 and R5) from high traffic loads generated by NCTs in the programmable logic (PL).
- NCT Performance Maintained: While isolating CTs, the QIQoS setups also ensured that NCTs still received service, preventing complete starvation and allowing for a balance between isolation and overall system throughput.
- Reduced Timing Uncertainty: The results indicate that QIQoS significantly narrows the gap between timing estimates obtained during early development (using synthetic aggressors) and actual system performance.
Limitations and Future Work
- Not Full Isolation: The authors acknowledge that the isolation is "quasi" and not absolute. There will always be some residual interference, especially under extreme conditions.
- Platform Specificity: The QIQoS setups are instantiated on the Xilinx Zynq UltraScale+ DDRMC. While the principles are generalizable, concrete implementations would require adaptation for other MPSoCs and QoS controllers.
- Future Work: The authors plan to explore other QoS features of the DDRMC to further isolate APU and RPU cores from PL traffic, and also to isolate cores within the APU and RPU from each other. They also aim to investigate how QIQoS can be applied to other shared resources beyond the DDRMC.
Summary updated