TL;DR
This paper proposes an ontology-based approach to bridge the gap between generic resource schedulers and application-specific knowledge, enabling more energy-efficient resource management. By modeling system states and application behaviors in OWL ontologies, they demonstrate that current reasoners can identify critical situations in near real-time, allowing for proactive adaptations.
What Problem Does This Paper Address?
Modern computing systems are increasingly complex and energy-hungry. While operating systems (OS) manage resources, they often lack the detailed knowledge of how applications actually utilize those resources. This "application knowledge gap" prevents generic resource schedulers from making optimal energy-saving decisions. For instance, a scheduler might over-provision resources simply because utilization is high, without understanding if the application can truly benefit from them or if a different configuration would be more efficient. This leads to wasted energy and potentially suboptimal performance. The paper aims to resolve this by enabling resource schedulers to leverage application-specific insights without becoming tightly coupled to individual applications.
Methods
The core of the proposed solution is an ontology-based situation recognition system.
- Ontology Modeling: They use OWL 2 EL (a lightweight ontology language) to model the system's hardware and software components, as well as application-specific knowledge. This TBox (Terminological Box) defines the "what" of the system.
- Situation Definition: Critical situations that warrant resource scheduling adaptations are defined as class queries or conjunctive queries over the ontology. These queries represent specific states or patterns that indicate a need for action.
- Runtime ABox: At runtime, the current state of the system and application information are captured in an ABox (Assertion Box). This ABox is dynamically updated.
- Ontology Reasoning: Standard OWL 2 EL reasoners are employed to query the ABox against the defined situations. If a query matches, the reasoner signals the resource manager.
- Bridging the Gap: This approach acts as a "glue layer," allowing the generic resource manager to receive actionable insights derived from application-specific knowledge without needing direct, application-specific interfaces.
Datasets
The paper does not use pre-existing datasets for training. Instead, they create a simulated environment based on a video platform use case.
- Ontology (OVP): A custom OWL 2 EL ontology was developed to represent the video platform's components, services, and application-specific knowledge. This ontology contains approximately 200 classes and 50 properties.
- ABoxes: Three distinct ABoxes were generated to represent different states of a mid-sized video platform running on 20 servers. These ABoxes contained a significant number of individuals, class assertions, and property assertions (e.g., over 1,000 individuals, 1,500 class assertions, and 104,000 property assertions per ABox).
Key Insights and Ideas
- Ontologies as a Unifying Language: OWL ontologies provide a formal and expressive way to integrate diverse information from different system layers and applications into a coherent model.
- Situation Recognition for Proactive Adaptation: By defining critical situations as queries, the system can proactively identify needs for resource adaptation (e.g., powering up servers before demand spikes) rather than reacting to already critical states.
- Decoupling Generic Schedulers: The ontology-based approach allows resource schedulers to remain generic, avoiding the need for application-specific code, thus promoting reusability and maintainability.
- Conjunctive Queries Offer More Expressivity: While class queries are simpler, conjunctive queries proved more effective for modeling certain situations, allowing for more complex relational patterns to be detected.
Key Results
The paper's experiments focused on evaluating the performance of state-of-the-art OWL 2 EL reasoners for situation recognition.
- Performance: For class queries, the fastest reasoners (e.g., ELK, FACT++, HermiT) completed the situation recognition task within 4 seconds, which is considered acceptable for mid-term adaptations.
- Conjunctive Queries: While more computationally intensive, conjunctive queries also yielded acceptable results, with TROWL completing the task within 2 seconds (for a subset of queries).
- Feasibility: The experiments demonstrated that OWL 2 EL can faithfully model the knowledge of a complex system like the video platform and that current reasoners can process realistic amounts of data within acceptable timeframes for situation recognition.
Limitations and Future Work
- Reasoner Soundness: The authors note that while TROWL showed excellent performance, its answers were not sound in all test cases, highlighting the need for careful reasoner selection.
- Temporal Information: The current approach does not explicitly handle temporal information. Future work aims to incorporate temporal logic (like LTL) to model situations more accurately, especially for real-world scenarios.
- Real-world Deployment: The current study is based on simulated data. Future work includes implementing the approach in a real-world setting to measure actual energy savings and impact on service quality.
- Cross-layer Integration: The authors envision extending the approach for cross-layer integration to resolve conflicting adaptations between different system layers.
Summary updated