The rapid expansion of digital services, particularly generative AI and cloud computing, has significantly increased the energy demands of data centers. This growth poses serious threats to both the economy and the environment by straining power grids, driving up electricity costs, increasing greenhouse gas emissions, and exacerbating water scarcity. While energy-aware computing practices have emerged as a promising solution, the role of operating system components, especially file systems remains underexplored. In this paper, we investigate the performance and energy efficiency implications of three widely used Linux file systems: ext4,$xfs$, and$f2fs$. Using a high-performance NVMe SSD, a power meter, and a diverse set of synthetic and real-world 10 workloads, we conduct a systematic evaluation across varying request sizes and concurrency levels. Our results reveal that$f2fs$excels in low-intensity random write workloads due to its log-structured design, while$xfs$consistently delivers the best energy efficiency under high-concurrency scenarios. We highlight how file system design choices can significantly influence system-wide energy consumption and advocate for energy-aware system software design as a critical step toward sustainable cloud infrastructure.

With the availability of high performance storage technology, there is extra pressure on the efficiency of IO interfaces. In addition to the popular POSIX synchronous, POSIX asynchronous, and Linux asynchronous (libaio) IO interfaces, there are two recent interfaces, spdk and io_uring, that are increasingly attracting attention with their high performance asynchronous designs. While providing high performance IO is crucial, it is also essential to do so in an energy-aware manner. In this paper, we study the energy implications of IO interface design choices and how these choices impact a system's energy consumption. Our empirical evaluation using a power meter, an ultra-low latency storage device, and various workload behaviors including single and multiple thread scenarios allow us to lay out the most energy efficient design choices, with the goal of yielding energy-aware high-performance storage stack designs.

The performance of recent data storage devices has significantly improved over previous generations, with lower latency, greater throughput, and greater parallelism. Since we now have Ultra-Low Latency (ULL) data storage devices capable of providing data in less than 10 microseconds, in this paper we question the need for IO schedulers for better performance and energy efficiency. Specifically, we measure the latency costs of Linux IO scheduling algorithms and investigate their impact on overall performance and energy efficiency using a ULL storage device, a power meter, and various IO workloads. Our observations indicate that IO schedulers for ULL storage either do not help or significantly increase request latencies while also negatively impacting throughput and energy efficiency. Although we recognize the value of IO schedulers for slower devices or for other metrics such as fairness and QoS, we believe that IO schedulers have become unnecessary for ULL devices to improve performance or energy efficiency.

Polling is commonly indicated to be a more suitable IO completion mechanism than interrupt for ultra-low latency storage devices. However, polling's impact on overall energy efficiency has not been thoroughly investigated. In this paper, contrary to common belief, we show that polling can also be more energy efficient than interrupt. To do so, we systematically investigate the energy efficiency of all available Linux IO completion mechanisms, including interrupt, classic polling, and hybrid polling using a real ultra-low latency storage device, a power meter, and various workload behaviors. Our experimental results indicate that although hybrid polling provides a good trade-off in CPU utilization, it is the least energy efficient, whereas classic polling is the most energy efficient for low latency IO requests. To the best of our knowledge, this is the first paper classifying polling as more energy efficient than interrupt for a real secondary storage device, and we hope that our observations will lead to more energy efficient IO completion mechanisms for new generation storage device characteristics.

Efficient and accurate detection of data access correlations can provide various storage system optimizations. However, existing offline detection techniques are costly in terms of computation and memory, and they rely on previously recorded disk I/O traces, thus wasting additional storage space and causing further I/O. Moreover, due to their offline nature, they are inadequate for allowing automatic optimizations. In this paper, we propose a real-time data access characterization framework that eliminates the drawbacks of offline analysis with a slight compromise in accuracy. The proposed framework continuously monitors I/O requests and forms correlations by applying single-pass data analysis techniques and maintaining a synopsis data structure to efficiently characterize access behavior in various dimensions including spatial locality, temporal locality, and frequency. Our evaluation using synthetic and real world storage workloads indicates that the proposed framework can detect over 90% of data access correlations in real-time, using limited memory. The proposed framework is crucial for enabling future self-optimizing storage systems that can automatically react to I/O bottlenecks.

The growing energy consumption of data centers is a compelling global problem and effective server consolidation is at the heart of energy efficient cloud data centers. A variant of bin packing can be used to model the server consolidation problem, where the constraints are multidimensional and heterogeneous vectors rather than scalars and the goal is to satisfy the requested resource allocation using the minimum number physical servers. Since bin packing is NP-hard, we rely on heuristics for practical solutions. Variations of First Fit Decreasing (FFD) based heuristics have been shown to be effective both in theory and practice for the one dimensional homogeneous case. However, the multidimensional and heterogeneous aspects of the server consolidation problem make it more complicated, requiring additional research to adapt FFD to the server consolidation problem. In this paper, we present a new FFD-based server consolidation technique using a Monte Carlo method and Shannon entropy, which considers resource bottlenecks and dynamically adjusts to variance in the utilization of different resources. The proposed heuristic outperforms existing techniques in all scenarios, achieving within 2-5% of optimal on average for medium to high variance in resource utilization, and within 10% worse than optimal on average for all scenarios.

While society continues to be transformed by insights from processing big data, the increasing rate at which this data is gathered is making processing in private clusters obsolete. A vast amount of big data already resides in the cloud, and cloud infrastructures provide a scalable platform for both the computational and I/O needs of big data processing applications. Virtualization is used as a base technology in the cloud; however, existing virtual machine placement techniques do not consider data replication and I/O bottlenecks of the infrastructure, yielding sub-optimal data retrieval times. This paper targets efficient big data processing in the cloud and proposes novel virtual machine placement techniques, which minimize data retrieval time by considering data replication, storage performance, and network bandwidth. We first present an integer-programming based optimal virtual machine placement algorithm and then propose two low cost data- and energy-aware virtual machine placement heuristics. Our proposed heuristics are compared with optimal and existing algorithms through extensive evaluation. Experimental results provide strong indications for the superiority of our proposed solutions in both performance and energy, and clearly outline the importance of big data aware virtual machine placement for efficient processing of large datasets in the cloud.

Storage performance bottlenecks are one of the major threats limiting the scalability of I/O intensive applications. Parallel storage systems have the potential to alleviate I/O bottlenecks through concurrent operation of independent storage components if a parallelism-aware data layout can be continuously guaranteed. Existing systems use one-layout-fits-all data placement strategy that frequently results in sub-optimal I/O parallelism. Guided by association rule mining, graph coloring, bin packing, and network flow techniques, this paper proposes a general framework for self-optimizing parallel storage systems, with the goal of continuously providing a high-degree of I/O parallelism that is robust to changes in the parallel access patterns of applications and the coexistence of applications with different parallel access characteristics. Evaluation results indicate that the proposed framework is highly successful in adjusting to skewed parallel access patterns for both traditional hard disk drive (HDD) based storage arrays and solid-state drive (SSD) based all-flash arrays. In addition to the storage arrays, the proposed framework is sufficiently generic to be tailored to various other parallel storage scenarios including but not limited to key-value stores, parallel/distributed file systems, and internal parallelism of SSDs.