Given the emerging needs of artificial intelligence in both commercial and consumer uses, the rapid increase of energy consumption has become a major concern for data centers. Workload scheduling plays a crucial role in reducing the energy costs associated with large compute-intensive workloads in data centers. However, there has not been much research on reducing the energy costs for GPU-intensive workloads specifically. In this paper, we proposed two thermal-ware scheduling algorithms (GPUThermalAwareMinTemp and GPUThermalAware) for GPU-intensive workloads. Through the integration of machine learning models into a modified version of the GPUCloudSimPlus simulator, we compare the performance of our algorithms with three baseline algorithms by simulating cloud computing environments using a subset of real-world workload traces from Alibaba PAI. Our experimental results show that, compared with the worst-case algorithms, the GPU thermal-aware scheduling algorithms achieved 4.72% lower average host minimum temperatures and saved total energy by 1.17%.

Dynamic thermal-aware load balancing addresses rising data center cooling costs by relying on accurate GPU temperature prediction. Due to the complexity of GPU temperature, deep learning models are widely used, but exposure bias–differences between training and inference environments–often degrades performance in autoregressive settings. While methods like scheduled sampling and attention forcing have been proposed to mitigate exposure bias, their effectiveness for hardware temperature forecasting is not well understood. This study systematically evaluates these mitigation techniques along with patching and modern deep learning architectures for GPU temperature prediction, considering accuracy, robustness, speed, and energy efficiency. Results show that scheduled sampling significantly worsens autoregressive RMSE and exposure bias (p < 0.001) for CNN-BiLSTM and Transformer models, whereas scheduled attention forcing reduces exposure bias but increases RMSE. The classic Transformer achieved the best overall balance with an AR RMSE of 4.34, median exposure bias of 2.4, median inference time of 257 µs, and the lowest training power usage (0.42 kWh), resulting in 22–84% lower carbon emissions. Patch-based models like PatchTST excelled in non-autoregressive settings but only the smallest patch sizes approached the Transformer’s autoregressive robustness at higher computational cost.

California’s Mediterranean climate, characterized by mild and wet winters and hot and dry summers, makes the state particularly susceptible to wildfires. Increasingly severe fires, exacerbated by climate change, require accurate predictions to assess the risk of their occurrence. Machine learning models have been used in the past for wildfire prediction, but have been computationally intensive as they have focused on California in its entirety. This project aims to reduce the computational cost by focusing on three distinct prediction zones in California: the Southern Mediterranean coast, the Central Sierra Nevada Mountains, and the Northern California coast. Additionally, the imbalanced fire and non-fire data has resulted in low precision and recall scores in the past. This paper aims to improve the performance, specifically precision and recall, of fire prediction models when working with inherently imbalanced data. The processing times of our models were all under an hour, with the fastest being 36 minutes, indicating a decrease in the computational efforts needed. The best performing model, the Central Zone, had an accuracy of 99.96% with a recall and precision of 87.66% and 96.35% respectively. These results indicate that by separating California into multiple homogeneous ecosystem-based prediction zones the computational efforts were limited while the model performance was improved. Future work in this project includes extending the analyses to more ecosystems in California.

Cooling costs count for a significant part of the total energy consumption in data centers, and previous researchers mainly focused on investigating thermal-ware workload distribution strategies for CPU-intensive workloads. This paper introduces a novel machine learning-based approach that aims at reducing energy consumption through thermal-aware workload distribution to build energy-efficient data centers for GPU-intensive workload. To achieve this goal, the study employs the GPUCloudSim Plus simulator, which effectively models the distribution of GPU-intensive applications under diverse workloads and utilizations. The integration of advanced machine learning models allows for accurate temperature predictions and comprehensive evaluation of the proposed algorithm's performance. We evaluated our ThermalAwareGpu workload scheduling algorithm, and it saved up to 12.82% of computing cost compared to the baseline algorithms. Our future work will explore the estimation of data center cooling energy and conduct in-depth comparisons of different workload balancing algorithms on more intensive experiments.

Certificate authentication is considered to be one of the most tedious and complex processes. Various types of documents need different authentication and verification methods. When students go through multiple phases of education, they are required to submit documentsfor admission, scholarship application, and tuition concessions. Hosting these certificates in a secure and tamper-proof manner on a decentralized blockchain will allows participating parties to view transparent certificates. In this paper, we proposed and implemented a Soulbound Token Certification (SBTCert) Verification System, a blockchain-based digital certificate platform, that aims to enhance the verification model without compromising security, authentication, decentralization, and confidentiality. The system provides a secure and reliable platform for universities, students, and staff to manage, issue, and verify educational certificates.

Vehicle detection and classification are important for intelligent transportation planning, directly impacting traffic flow and management efficiency. There have been a lot of research on vehicle detection and classification for autonomous driving, but research on roadside traffic flow has not been fully investigated. This research explores various methods for vehicle detection and classification that could be applied to 3D roadside traffic flow analysis. Using data collected from a Velodyne 32C Lidar sensor 3D camera and a stereo-based depth 2D camera, this project implemented a Complex-YOLO projection-based model, by converting 3D point cloud data into a 2D bird-eye-view projection and applying the YOLO model for vehicle detection and classification. Our experimental results demonstrated that using transfer learning as a training technique, along with normalization of the rotation angle, enhanced the model performance for vehicle detection and classification. Without normalization of the rotation angle, the highest precision achieved was 89.71 %, with a recall of 85.15 %. With normalization of the rotation angle, the best model achieved a precision of 92.20 % with a recall of 84.90 %, demonstrating the effectiveness of both transfer learning and normalization in handling and aligning the dataset for more effective learning.

Energy-efficient workload management is increasingly important in data centers as demand for cloud computing continues to grow. Recent findings in workload management point to thermal-aware load balancers as an effective approach to reduce energy consumption via cooling costs. Due to the high energy cost of Data Centers (DCs) and possible risk of disrupting critical services, simulations can be used to evaluate temperature-aware load-balancing algorithms. In this paper, we leverage an existing DC simulator's environmental parameters for integrating a temperature prediction model to construct a temperature-aware workload distribution simulation. Using scheduled sampling, we implemented a custom recursive multi-step forecasting model that reduced the discrepancy between training error and simulation predictions. Compared with a baseline GPU temperature prediction model, our hybrid CNN-LSTM model achieved a reduction in RMSE from 2.54 to 0.73.

Energy-efficient workload management has been widely adopted by data centers for cloud computing. Thermal and energy modeling plays an important role in making decisions on workload management. Machine learning technology has become increasingly popular used in thermal modeling. In this paper, we studied existing machine learning algorithms and methods for thermal prediction for data centers and conducted experiments to investigate the impact of activities on the temperature and energy consumption with CPU-intensive workload. We collected the CPU utilization, temperature, and energy data, and applied several regression models and the XG Boost machine learning model to predict the temperature of the CPU. Performance of the regression models was compared with the XGBoost machine learning models. With more experiments are been conducting to investigate the CPU temperature under various combinations of CPU core utilizations, we will further improve the performance of the XGBoost machine learning model.

California continuously ranks as the most wildfire prone state in the United States, with 19 billion dollars in fire damage from 2020 alone. The scarcity of wildfire prediction research specifically for California only increases the state's vulnerability to increasingly frequent and severe wildfires. In this paper, we applied four machine learning models to make wildfire risk predictions from NASA remote sensing data and explored the application of 3D computer graphics software to visualize ordinarily complex geospatial information. From 2D satellite images, we were able to achieve unique 3D visualizations and animations which allowed viewers to easily digest the significance of each respective geospatial dataset over a period of 10 years. Among the four machine learning models, the Multi-Layer Perceptron (MLP) model was the most effective in predicting wildfire risk with an 82.97% accuracy. Although accurate, the imbalance of data between fire and non fire data points impacted the precision of our predictions, warranting further consideration. This imbalance highlights an area for improvement in future research efforts and underscores the need for a more comprehensive and balanced dataset to enhance the precision of our wildfire risk predictions.

Consumer interest in cosmetics, and particularly skincare products, has surged globally in recent years. Traditional methods of selecting skincare products involve relying on best-sellers or in-store recommendations. However, these approaches are ineffective because they fail to account for individual variations in skin conditions and consumer compatibility. This research aims to design a skincare product recommendation system based on users' skin types and ingredient compositions of products. The proposed method employs content-based filtering to identify chemical components of products and find products with similar ingredient compositions. Unlike many existing sys-tems that require users to input product names, the new system takes into account their desired beauty effects to accommodate those with limited skincare knowledge. The resulting system returns personalized recommendations across multiple product categories. It could contribute to improved compatibility between users and skincare products, enhanced user satisfaction and streamlined selection processes.

Cooling costs count for a significant part of the total energy consumption in data centers, and previous re-searchers mainly focused on investigating thermal-ware workload distribution strategies for CPU-intensive workloads. This paper introduces a novel machine learning-based approach that aims at reducing energy consumption through thermal-aware workload distribution to build energy-efficient data centers for GPU-intensive workload. To achieve this goal, the study employs the GpuCloudSim Plus simulator, which effectively models the dis-tribution of GPU-intensive applications under diverse workloads and utilizations. The integration of machine learning models allows for accurate temperature predictions and comprehensive evaluation of the proposed algorithm's performance. We pro-posed a new workload scheduling algorithm, ThermalAwareGpu, to reduce the energy cost for GPU-intensive workload. We evaluated our algorithm by generating three common patterns of workloads, and saved up to 12.79 % of computing cost compared to the baseline algorithms. Our future work includes exploring the estimation of data center cooling energy and conducting in-depth comparisons of different workload balancing algorithms on various compute-intensive workloads.

Wildfires in California are increasingly frequent and caused a vast range of damages to the land and society. Existing work in the field of machine learning aided fire susceptibility mapping contains promising results, however they focused on small fire prone areas or yielded less than impressive classification metrics, with even fewer groups focusing on California. In this paper, we applied five machine learning models for fire prediction using a series of remote sensing data and fire incident records of California. The fire incident records were used for labeling fire conditions, and remote sensing features (land surface temperature, normalized difference vegetation index, and thermal anomalies) were selected for model training. Two datasets were created for analysis of wildfires in California based on different sized regions: the whole state treated as one region, and California being broken into counties. For the state level dataset, fire prediction accuracies were 89% for artificial neural network (ANN), and 87% for Gaussian Naive Bayes (GaussianNB), k-nearest neighbors (KNN), and logistic regression (LR), and support vector machine (SVM). For the county level dataset, the prediction accuracies were 97% for KNN, 96% for ANN, SVM, and LR, and 94% for GuassianNB.

Data caching strategies and Garbage Collection on SSDs have been extensively explored in the past years. However, the Mapping Table cache performance has not been well studied. Mapping table provides page translation information to Flash Translation Layer (FTL) in order to translate Logical Page Address (LPA) to Physical Page Address (PPA). Missing in mapping table cache causes extra read transactions to flash storage which results in stalls of I/O requests processing in SSDs. Random read requests are affected more than random write requests since write requests can be handled by write cache effectively. In this paper, we analyze the impact of CMT on different random read requests and present a Cached Mapping Table prefetching approach which fetches logical-to-physical page translation information in order to mitigate the stalls in processing random read requests. Our experimental results show an improvement of average request waiting time by up to 13%.

This research proposes an advanced approach that improves the efficiency of skyline query processing by significantly reducing the computational cost on object comparisons, i.e., dominance tests between objects. Our solutions are based on two novel concepts. The skyline query over Minimum Bounding Rectangles (MBRs) receives a set of MBRs and returns the MBRs that are not dominated by other MBRs. In the dominance test for MBRs, the detailed attribute values of objects in the MBRs are not accessed. Moreover, the dependent group of MBRs reduces the search space for dominance tests. Objects in an MBR are only compared with the ones in the corresponding dependent groups of the MBR rather than with the entire dataset. Our solutions apply the two concepts to the R-tree in order to use its hierarchical structure in which every node is a natural abstraction of an MBR. Specifically, given the R-tree index of an input dataset, we first eliminate unqualified objects by utilizing the skyline query over MBRs (i.e., intermediate nodes in the R-tree). Subsequently, we generate dependent groups for the skyline MBRs. Two dependent group generation methods that rely on either the sorting technique or the R-tree index are developed. Further, we apply an existing skyline algorithm to every dependent group, and the results of the original skyline query are the union of skyline objects in the dependent groups. In addition, we also analyze the cardinality of the two new concepts based on a probabilistic model, which enables us to analyze the computational complexity of the proposed solutions. Our experimental results show that the proposed solutions are clearly more efficient than the state-of-the-art approaches.

This paper presents an exact solution and a heuristic solution to a UAV-assisted parcel delivery problem, in which UAVs can only be operated in Visual-Line-Of-Sight (VLOS) areas. In our proposed problem, we assume that trucks travel on road networks, and UAVs move in Euclidean spaces and can launch at any locations on roads. We first demonstrate the overview of our exact solution that iterates all permutations of destinations for an optimal delivery route. Given a specific delivery order, an intuitive approach needs to check all possible locations on roads in the VLOS areas and find a globally optimal location for every destination if UAVs are used for delivery. To avoid high computational cost of searching the optimal location at runtime, we propose an advanced index-based alternative, which computes optimal delivery routes in a pre-processing stage. Due to the nature of NP-hard problems, we also propose a heuristic approach that utilizes delivery groups for the proposed problem of practical size. All proposed solutions are evaluated through extensive experiments.

In this paper, we present a thermal-aware file assignment technique called TIGER for reducing the cooling cost of storage clusters in data centers. We show that peak inlet temperatures of storage nodes depend on not only CPU utilization but also I/O activities, which rely on file assignments in a cluster. The TIGER scheme aims to lower peak inlet temperatures of storage clusters by dynamic thermal management through file placements. TIGER makes use of cross-interference coefficients to estimate the re-circulation of hot air from the outlets to the inlets of data nodes. TIGER first calculates the thresholds of disks in each data node based on its contribution to heat re-circulation in a data center. TIGER undertakes two steps to achieve high I/O performance while reducing cooling cost. First, TIGER assigns groups of files with similar service times to shorten I/O response times. Second, TIGER ensures that load imbalance does not exceed a specified threshold. We evaluate performance of TIGER in terms of both cooling energy conservation and response time of a storage cluster. Our results confirm that TIGER reduces cooling-power requirements for clusters by offering about 10 to 15 percent cooling-energy savings without significantly degrading I/O performance.

This research develops an advanced two-phase MapReduce solution that is able to efficiently address skyline queries on large datasets. Unlike existing parallel skyline approaches, our scheme considers data partitioning, filtering, and parallel skyline evaluation as a holistic query process. In particular, we apply filtering techniques and angle-based partitioning in the first phase, in which unqualified objects are discarded and the processed objects are partitioned by their angles to the origin.In the second phase, local skyline objects in each partition are calculated in parallel, and global skyline objects are output after a merging skyline process. To improve the parallel local skyline calculation, we propose two partition-aware filtering methods that keep skyline candidates in a balanced manner. The aggressive partition-aware filtering aggressively eliminates objects in the partition with the greatest population of candidate objects, whereas the proportional partition-aware filtering slows down the growth of partition population proportionally. Recognizing the lack of studies that incorporate the MapReduce framework into parallel skyline processing, we propose a partialpresort grid-based partition skyline algorithm that is able to significantly improve the merging skyline computation on large datasets. The presort process can be completed in the shuffle phase with little overhead. Our experimental results show the efficiency and effectiveness of the proposed parallel skyline solution utilizing MapReduce on large-scale datasets.

In this demonstration, we present a visualization system offering two advanced solutions that can efficiently address a novel multi-criteria optimal location query by using Overlapping Voronoi Diagrams (OVDs). Our system not only displays an example that applies the advanced solutions to a practical optimal location query, but also visualizes the process of query evaluation by using the solutions, which can help users better understand the two solutions in a more intuitive way. The Real Region as Boundary (RRB) approach calculates real regions of overlapped Voronoi cells, while the Minimum Bounding Rectangle as Boundary (MBRB) approach approximates the cells by their MBRs to simplify region overlapping calculation. For efficient query evaluation and better demonstration, global and local datasets are managed separately and only the local dataset is displayed in a limited visualization area. User interactions are provided in diverse manners in our system by using the Standard Widget Toolkit (SWT) and Open Graphics Library (OpenGL).

This article presents a framework, Frog, for Context-Based File Systems (CBFSs) that aim at simplifying the development of context-based file systems and applications. Unlike existing informed-based context-aware systems, Frog is a unifying informed-based framework that abstracts context-specific solutions as views, allowing applications to make view selections according to application behaviors. The framework can not only eliminate overheads induced by traditional context analysis, but also simplify the interactions between the context-based file systems and applications. Rather than propagating data through solution-specific interfaces, views in Frog can be selected by inserting their names in file path strings. With Frog in place, programmers can migrate an application from one solution to another by switching among views rather than changing programming interfaces. Since the data consistency issues are automatically enforced by the framework, file-system developers can focus their attention on context-specific solutions. We implement two prototypes to demonstrate the strengths and overheads of our design. Inspired by an observation that there are more than 50% of small files (<4KB) in a file system, we create a Bi-context Archiving Virtual File System (BAVFS) that utilizes conservative and aggressive prefetching for the contexts of random and sequential reads. To improve the performance of random read-and-write operations, the Bi-context Hybrid Virtual File System (BHVFS) combines the update-in-place and update-out-of-place solutions for read-intensive and write-intensive contexts. Our experimental results show that the benefits of Frog-based CBFSs outweigh the overheads introduced by integrating multiple context-specific solutions.

An increasing number of spatial keyword query techniques that return qualified objects based on a comprehensive consideration of spatial and keyword constraints have been presented in the literature. Due to the complexity of the solutions to spatial keyword queries, systems that can effectively demonstrate the mechanisms will attract interest from the spatial database research community. However, very limited visualization systems have been developed for illustrating spatial keyword query evaluation. In this demonstration, we present a system that visualizes advanced solutions to efficiently answer the Spatial Keyword k Nearest Neighbor (SKkNN) query. With the two-level data management method and the friendly user interface implemented by the Standard Widget Toolkit (SWT) and Open Graphics Library (OpenGL), our system is able to not only interact with users in diverse manners, visualize datasets, and display the SKkNN query evaluation process, but it also helps users better understand the solutions in a more intuitive way.

Numerous geographic information system applications need to retrieve spatial objects which bear user specified keywords close to a given location. In this research, we present efficient approaches to answer spatial keyword queries on spatial networks. In particular, we formally introduce definitions of Spatial Keyword k Nearest Neighbor (SKkNN) and Spatial Keyword Range (SKR) queries. Then, we present a framework of a spatial keyword query evaluation system which is comprised of Keyword Constraint Filter (KCF), Keyword and Spatial Refinement (KSR), and the spatial keyword ranker. KCF employs an inverted index to calculate keyword relevancy of spatial objects, and KSR refines intermediate results by considering both spatial and keyword constraints with the spatial keyword ranker. In addition, we design novel algorithms for evaluating SKkNN and SKR queries. These algorithms employ the inverted index technique, shortest path search algorithms, and network Voronoi diagrams. Our extensive simulations show that the proposed SKkNN and SKR algorithms can answer spatial keyword queries effectively and efficiently.

In this paper, we present thermal-aware file assignment technique called TIGER for reducing cooling cost of storage clusters in data centers. TIGER first calculates the thresholds of disks in each node based on its contribution to heat recirculation in a data center. Next, TIGER assigns files to data nodes according to calculated thresholds. We evaluated performance of TIGER in terms of both cooling energy conservation and response time of a storage cluster. Our results confirm that TIGER reduces cooling-power requirements for clusters by offering about 10 to 15 percent cooling energy savings without significantly degrading I/O performance.

This paper presents a novel Predictive Energy-Aware Management (PEAM) system that is able to reduce the energy costs of storage systems by appropriately selecting data transmission methods. In particular, we evaluate the energy costs of three methods (1. transfer data without archiving and compression, 2. archive and transfer data, 3. compress and transfer data) in preliminary experiments. According to the results, we observe that the energy consumption of data transmission greatly varies case by case. We cannot simply apply one method in all cases. Therefore, we design an energy prediction model that can estimate the total energy cost of data transmission by using particular transmission methods. Based on the model, our predictive energy-aware management system can automatically select the most energy efficient method for data transmission. Our experimental results show that our system performs better than simply selecting any one among the three methods for data transmission in terms of energy efficiency.

We present a multicore-enabled smart storage for clusters in general and MapReduce clusters in particular. The goal of this research is to improve performance of data-intensive parallel applications on clusters by offloading data processing to multicore processors in storage nodes. Compared with traditional storage devices, next-generation disks will have computing capability to reduce computational load of host processors or CPUs. With the advance of processor and memory technologies, smart storage systems are promising devices to perform complex on-disk operations. The proposed smart storage system can avoid moving a huge amount of data back and forth between storage nodes and computing nodes in a cluster. To enhance the performance of data-intensive applications, we have designed a smart storage system called Multicore-enabled Smart Storage (McSD), in which a multicore processor is integrated in storage nodes. We have implemented a programming framework for data-intensive applications running on a computing system coupled with McSD. The programming framework aims at balancing load between computing nodes and multicore-enabled smart storage nodes. To fully utilize multicore processors in smart storage nodes, we have implemented the MapReduce model for McSDs to handle parallel computing on a cluster. A prototype of McSD has been implemented in a cluster connected by Gigabit Ethernet. Experimental results show that McSD can significantly reduce the execution times of three real-world applications - word count, string matching, and matrix multiplication. We demonstrate that the integration of multicore-enabled smart storage with MapReduce clusters is a promising approach to improving overall performance of data-intensive applications on clusters.

Recognizing that power and cooling cost for data centers are increasing, we address in this study the thermal impact of storage systems. In the first phase of this work, we generate the thermal profile of a storage server containing three hard disks. The profiling results show that disks have comparable thermal impacts as processing and networking elements to overall storage node temperature. We develop a thermal model to estimate the outlet temperature of a storage server based on processor and disk utilizations. The thermal model is validated against data acquired by an infrared thermometer as well as build-in temperature sensors on disks. Next, we apply the thermal model to investigate the thermal impact of workload management on storage systems. Our study suggests that disk-aware thermal management techniques have significant impacts on reducing cooling cost of storage systems. We further show that this work can be extended to analysis the cooling cost of data centers with massive storage capacity.

Most researches of Solid State Drives (SSDs) architectures rely on Flash Translation Layer (FTL) algorithms and wear-leveling; however, internal parallelism in Solid State Drives has not been well explored. In this research, we proposed a new strategy to improve SSD write performance by enhancing internal parallelism inside SSDs. A SDRAM buffer is added in the design for buffering and scheduling write requests. Because the same logical block numbers may be translated to different physical numbers at different times in FTL, the on-board SDRAM buffer is used to buffer requests at the lower level of FTL. When the buffer is full, same amount of data will be assigned to each storage package in SSDs to enhance internal parallelism. To accurately evaluate performance, we use both synthetic workloads and real-world applications in experiments. We compare the enhanced internal parallelism scheme with the traditional LRU strategy since it is unfair to compare an SSD having buffer with an SSD without a buffer. The simulation results demonstrate that the writing performance of our design is significantly improved compared with the LRU-cache strategy with the same amount of buffer sizes.

This paper presents an offloading framework - ORCA - to map I/O-intensive code to a cluster that consists of computing and storage nodes. To reduce data transmission among computing and storage nodes. our offloading framework partitions and schedules CPU-bound and I/O-bound modules to computing nodes and active storage nodes, respectively. From developer's perspective, ORCA helps them to deal with execution-path control, offloading executable code, and data sharing over a network. Powered by the offloading APIs, developers without any I/O offloading or network programming experience are allowed to write new I/O-intensive code running efficiently on clusters. We implement the ORCA framework on a cluster to quantitatively evaluate performance improvements offered by our approach. We run five real-world applications on both homogeneous and heterogeneous computing environments. Experimental results show ORCA speeds up the performance of all the five tested applications by a factor of up to 90.1% with an average of 75.5%. Moreover, the results confirm that ORCA reduces network burden imposed by I/O-intensive applications by a factor of anywhere between 35 to 68.

In this paper, we present an informed prefetching technique called IPODS that makes use of application-disclosed access patterns to prefetch hinted blocks in distributed multi-level storage systems. We develop a prefetching pipeline in IPODS, where an informed prefetching process is divided into a set of independent prefetching steps among multiple storage levels in a distributed system. In the IPODS system, while data blocks are prefetched from hard disks to memory buffers in remote storage servers, data blocks buffered in the servers are prefetched through networks to clients' local cache. We show that these two prefetching steps can be handled in a pipelining manner to improve I/O performance of distributed storage systems. Our IPODS technique differs itself from existing prefetching schemes in two ways. First, IPODS reduces applications' I/O stalls by keeping hinted data in clients' local caches and storage servers' fast buffers (e.g., solid state disks). Second, in a prefetching pipeline, multiple informed prefetching mechanisms semi-dependently coordinate to fetch blocks (1) from low-level (slow) to high-level (fast) storage devices in servers and (2) from high-level devices in servers to clients' local cache. The prefetching pipeline in IPODS judiciously hides network latencies in distributed storage systems, thereby reducing the overall I/O access time in distributed systems. Using a wide range of real-world I/O traces, our experiments show that IPODS can improve noticeably I/O performance of distributed storage systems.