Professor of Computer Science, Onur Mutlu, has been awarded the Huawei OlympusMons Prize for his groundbreaking research on data storage technology and NAND circuits.
The OlympusMons Awards were introduced in 2019 to honor and support pioneering research on data storage and to promote collaboration between industry, academia, and research. In 2023, Professor Onur Mutlu’s team at ETH Zurich was recognized for their research on systems integrating data storage, networks, and computing. The 1 million Chinese yuan (137,883 USD) award was given for the development of an adaptive optimization algorithm. Their 2023 documentation, titled “Memory and Storage Systems Driven by Machine Learning,” presented three elements of the system’s operation:
1. Pythia – a self-healing memory controller using reinforcement learning.
2. Hermes – scientifically-based off-chip load prediction to accelerate long-latency requests.
3. Sibyl – self-healing controllers of a hybrid storage system (HSS) using reinforcement learning.
The idea behind using Sibyl and reinforcement learning is that a data placement controller operating at the storage matrix or operating system level can receive feedback on data placement decisions across two or three storage layers to minimize access time. Based on this information, it can optimize its data placement decisions, reducing overall access time and continuously learning to improve its performance.
Sibyl is detailed in a scientific publication from May 2022, titled “Sibyl: Adaptive and Scalable Data Placement in Hybrid Storage Systems using Online Reinforcement Learning.” The publication’s abstract states: “Sibyl analyzes various workload and storage device characteristics to make data placement decisions that take into account system characteristics. For each decision made, Sibyl receives a reward from the system, which it uses to evaluate the long-term impact of its decision on performance and continuously optimize its online data placement policy.”
Decisions take into account system-wide costs associated with data migrations between different storage layers.
Professor Mutlu’s team evaluated the effectiveness of Sibyl compared to data placement technologies based on CDE, HPS, Archivist, and RNN. According to the publication, Sibyl improves performance by 21.6% compared to the previous best data placement technique in a double HSS configuration. In a Sibyl configuration with three HSS – Optane P4800X SSD, Intel D3-S4510 SATA SSD, Seagate Barracuda 1TB SATA disk drive – performance increased by 48.2% compared to the state-of-the-art technique.
The system achieves 80% of the performance of an ideal data placement policy with complete knowledge of future access patterns, with a small storage overhead of only 124.4 KiB.
The publication’s conclusions state, “Our extensive evaluation of the system in a real-world environment demonstrates that Sibyl provides adaptability and scalability through continuous learning and adaptation to workload characteristics, storage configuration, and device characteristics, as well as providing system-level feedback to maximize the overall long-term performance of hybrid storage systems.”
Sibyl can be extended to incorporate new device layers and theoretically be applied to storage arrays in a hierarchical storage management system, although this has not yet been encoded and tested.
The source code for Sibyl, as well as Hermes and Pythia, are available on GitHub.
Frequently Asked Questions (FAQs) based on the main topics and information presented in the article:
1. Who received the Huawei OlympusMons award for research on data storage technology and NAND circuits?
Professor of Computer Science Onur Mutlu received the Huawei OlympusMons award.
2. When were the OlympusMons Awards introduced, and what is their purpose?
The OlympusMons Awards were introduced in 2019 to honor and support pioneering research on data storage and to promote collaboration between industry, academia, and research.
3. What did Professor Onur Mutlu receive the award for in 2023?
Professor Onur Mutlu received the award for his research on systems integrating data storage, networks, and computing.
4. What are the three elements of the system’s operation presented in the 2023 documentation?
The three elements of the system’s operation are: Pythia – a self-healing memory controller using reinforcement learning, Hermes – scientifically-based off-chip load prediction to accelerate long-latency requests, and Sibyl – self-healing controllers of a hybrid storage system (HSS) using reinforcement learning.
5. What is the idea behind using Sibyl with reinforcement learning?
The idea behind using Sibyl with reinforcement learning is that a data placement controller at the storage matrix or operating system level receives feedback on data placement decisions across different storage layers to minimize access time. Based on this information, the controller optimizes its data placement decisions, reducing overall access time and improving performance.
6. How does Sibyl analyze the workload and storage device characteristics?
Sibyl analyzes various workload and storage device characteristics to make data placement decisions considering the system’s characteristics.
7. How was Sibyl evaluated compared to other data placement technologies?
Professor Mutlu’s team evaluated the effectiveness of Sibyl compared to data placement technologies based on CDE, HPS, Archivist, and RNN. According to the publication, Sibyl improves performance by 21.6% compared to the previous best data placement technique in a double HSS configuration, and by 48.2% in a configuration with three HSS.
8. What is the conclusion of the publication regarding the Sibyl system?
The conclusion of the publication states that Sibyl provides adaptability and scalability through continuous learning and adaptation to workload characteristics, storage configuration, and device characteristics, by providing system-level feedback to maximize the overall long-term performance of hybrid storage systems.
9. Can Sibyl be expanded with new device layers?
Yes, Sibyl can be expanded with new device layers, and theoretically, it can be applied to storage arrays in a hierarchical storage management system, although this has not yet been encoded and tested.
10. Where can the source code for Sibyl and other components be found?
The source code for Sibyl, Hermes, and Pythia is available on GitHub.
Recommended related links to the main domain (not subpages) in [link name] format (if you are confident that the URL is 100% valid, do not add links to example.com):
– [ETH Zurich]
– [GitHub]
The source of the article is from the blog cheap-sound.com