Modern data-intensive systems rely on in-memory caching to achieve high throughput and low latency. CacheLib, Meta's general-purpose caching engine, provides high performance and flexibility for building specialized caches for a variety of applications. However, despite its wide adoption in large-scale infrastructures, CacheLib's data management mechanisms exhibit inefficiencies in shared environments. Particularly, its static and uncoordinated memory allocation leads to fragmented resource usage, unfair memory distribution, and degraded performance across tenants and instances.

DNN model versions are used for various tasks such as fine-tuning for downstream tasks, explainability, and debugging. Numerous checkpointing solutions exist that can be adapted to persist intermediate versions of a model, as it is being trained, at different storage locations. Additionally, version management tools allow us to log, visualize, compare, and query metadata related to ML, tracking changes made to previously built models. However, the version creation process of existing methods incurs high runtime and storage overheads. In this paper, we introduce LATTICE, a low-latency, direct persistence-based DNN versioning library for Non-Volatile Memory (NVM) expansion devices. LATTICE minimizes stalls during model versioning and reduces end-to-end versioning time by reorganizing the version creation workflow, streamlining memory allocation and deallocation for efficient snapshot creation, and leveraging multi-threaded parallelism. We also develop a user-friendly versioning API that transparently implements direct persistence. Our comprehensive evaluation with diverse DNN models shows that LATTICE can reduce persistence time by as much as 99.99%, decrease end-to-end versioning time by up to 72%, reduce versioning stalls by up to 35%, and increase versioning frequency by 0.2×-3.84× compared to state-of-the-art solutions. LATTICE also reduces space utilization for different workloads. The space savings are from 23.8% to 43.2% for workloads where model layers are progressively frozen and from 84.8% to 98.9% for fine-tuning workloads where only the last layers are tuned.

Graph Neural Networks (GNNs) have demonstrated remarkable capabilities across various domains. Despite the successes of GNN deployment, their utilization often reflects societal biases, which critically hinder their adoption in high-stake decision-making scenarios such as online clinical diagnosis, financial crediting, etc. Numerous efforts have been made to develop fair GNNs but they typically concentrate on either individual or group fairness, overlooking the intricate interplay between the two, resulting in the enhancement of one, usually at the cost of the other. In addition, existing individual fairness approaches using a ranking perspective fail to identify discrimination in the ranking. This paper introduces two innovative notions dealing with individual graph fairness and group-aware individual graph fairness, aiming to more accurately measure individual and group biases. Our Group Equality Individual Fairness (GEIF) framework is designed to achieve individual fairness while equalizing the level of individual fairness among subgroups. Preliminary experiments on several real-world graph datasets demonstrate that GEIF outperforms state-of-the-art methods by a significant margin in terms of individual fairness, group fairness, and utility performance.

Existing tiered memory systems all use DRAM-Preferred as their allocation policy, whereby pages get allocated from higher-performing DRAM until it is filled, after which all future allocations are made from lower-performing persistent memory (PM). The novel insight of this work is that the right page allocation policy for a workload can help to lower the access latencies for the newly allocated pages. We design, implement, and evaluate three page allocation policies within the real system deployment of the state-of-the-art dynamic tiering system. We observe that the right page allocation policy can improve the performance of a tiered memory system by as much as 17x for certain workloads.

The need to support new features in existing storage systems is an ongoing concern for storage developers. So is the desire to develop next generation storage systems that can adopt newly developed feature improvements with relative ease. Extending storage systems is challenging because of the inherent complexity of their codebases and the need to ensure that the storage state does not become corrupt or inconsistent when enabling new features. In this work, we examine a new storage architecture, FDMI, that uses the well-established publish-subscribe model for extending the feature set of a host storage system using plugins. A central mechanism in FDMI is transactional coupling. With transactional coupling, the subscribed plugin can either create new transactions that execute asynchronously following the successful completion of the precipitating event or can participate in the pending transaction and control whether the precipitating event itself will or will not be committed. We further create a classification of transactional mechanisms as well as possible desired plugin functionality and explore the matrix of these two classifications to create a new model for faster, safer distributed storage development.

The rapid growth of i-memory computing powered by data-intensive applications has increased demand for DRAM in servers. However, a DRAM-based system can be limiting for modern workloads because of its capacity, cost, and power consumption characteristics. Hybrid memory systems, which consist of different types of memory, such as DRAM and persistent memory, can help address many of these limitations. One promising direction that has been explored in the recent literature involves introducing persistent memory devices as a second memory tier that is directly exposed to the CPU. The resulting tiered memory design must address the fundamental challenge of placing the right data in the right memory tier at the right time while minimizing overhead. We present MULTI -CLOCK, an efficient, low-overhead hybrid memory system that relies on a unique page selection technique for tier placement. MULTl-CLOCK’s page selection captures both page access recency and frequency, and enables moving pages to appropriate tiers at the right time within hybrid memory systems. We implemented a Linux-based, NUMA-aware version of MULTI-CLOCK that is entirely transparent and backward compatible with any existing application. Our evaluation with diverse real-world applications such as graph processing and key-value stores shows that MULTI -CLOCK can improve the average throughput by as much as 352% when compared with several state-of-the-art techniques for tiered memory.

Storage is the Achilles heel of hybrid cloud deployments of workloads. Accessing persistent state over a WAN link, even a dedicated one, delivers an over-whelming performance blow to application performance. We propose FAB, a new storage architecture for the hybrid cloud. FAB addresses two major challenges for hybrid cloud storage, performance efficiency and backup efficiency. It does so by creating a new FAB layer in the storage stack that enables fault-tolerance, performance acceleration, and backup for FAB storage volumes. A preliminary evaluation of FAB's performance acceleration mechanism when deployed over Ceph's distributed block storage system offers encouragement to pursue this new hybrid cloud storage architecture.

Data centers today host large numbers of workloads and many of these workloads consume significant storage resources. Given the long history of successes in storage caching, it is only natural such successes bear fruit in modern data centers, at scale. This paper presents CaaS, a generalized caching service for cloud data centers. Departing from existing application, storage, or data-type specific caches, CaaS unifies and abstracts data center caching resources making these available to any workload and for any data type. Also departing from past caching practices, CaaS is fault-tolerant allowing it to cache writes without risk of data loss. We expect that systems such as CaaS will help bridge the gap between heterogeneous and distributed cache resources and data-intensive applications in a data center.

Efficient system-wide memory management is an important challenge for over-commitment based hosting in virtualized systems. Due to the limitation of memory domains considered for sharing, current deduplication solutions simply cannot achieve system-wide deduplication. Popular memory management techniques like sharing and ballooning enable important memory usage optimizations individually. However, they do not complement each other and, in fact, may degrade individual benefits when combined. We propose $\mathsf{Synergy}$Synergy, a hypervisor managed caching system to improve memory efficiency in over-commitment scenarios. $\mathsf{Synergy}$Synergy builds on an exclusive caching framework to achieve, for the first time, system-wide memory deduplication. $\mathsf{Synergy}$Synergy also enables the co-existence of the mutually agnostic ballooning and sharing techniques within hypervisor managed systems. Finally, $\mathsf{Synergy}$Synergy implements a novel file-level eviction policy that prevents hypervisor caching benefits from being squandered away due to partial cache hits. $\mathsf{Synergy}$Synergy's cache is flexible with configuration knobs for cache sizing and data storage options, and a utility-based cache partitioning scheme. Our evaluation shows that $\mathsf{Synergy}$Synergy consistently uses 10 to 75 percent lesser memory by exploiting system-wide deduplication as compared to inclusive caching techniques and achieves application speedup of 2x to 23x. We also demonstrate the capabilities of $\mathsf{Synergy}$Synergy to increase VM packing density and support for dynamic reconfiguration of cache partitioning policies.

No abstract available.

Persistent Memory devices present properties that are uniquely different from prior technologies for which applications have been built. Unfortunately, the conventional approach to building applications fail to either efficiently utilize these new devices or provide programmers a seamless development experience. We have built L ib PM, a Persistent Memory Library that implements an easy-to-use container abstraction for consuming PM. LibPM’s containers are data hosting units that can store arbitrarily complex data types while preserving their integrity and consistency. Consequently, L ib PM’s containers provide a generic interface to applications, allowing applications to store and manipulate arbitrarily structured data with strong durability and consistency properties, all without having to navigate all the myriad pitfalls of programming PM directly. By providing a simple and high-performing transactional update mechanism, L ib PM allows applications to manipulate persistent data at the speed of memory. The container abstraction and automatic persistent data discovery mechanisms within L ib PM also simplify porting legacy applications to PM. From a performance perspective, L ib PM closely matches and often exceeds the performance of state-of-the-art application libraries for PM. For instance, L ib PM ’s performance is 195× better for write intensive workloads and 2.6× better for read intensive workloads when compared with the state-of-the-art P mem .IO persistent memory library.

Defenders of enterprise networks have a critical need to quickly identify the root causes of malware and data leakage. Increasingly, USB storage devices are the media of choice for data exfiltration, malware propagation, and even cyber-warfare. We observe that a critical aspect of explaining and preventing such attacks is understanding the provenance of data (i.e., the lineage of data from its creation to current state) on USB devices as a means of ensuring their safe usage. Unfortunately, provenance tracking is not offered by even sophisticated modern devices. This work presents ProvUSB, an architecture for fine-grained provenance collection and tracking on smart USB devices. ProvUSB maintains data provenance by recording reads and writes at the block layer and reliably identifying hosts editing those blocks through attestation over the USB channel. Our evaluation finds that ProvUSB imposes a one-time 850 ms overhead during USB enumeration, but approaches nearly-bare-metal runtime performance (90% of throughput) on larger files during normal execution, and less than 0.1% storage overhead for provenance in real-world workloads. ProvUSB thus provides essential new techniques in the defense of computer systems and USB storage devices.

With the growing adoption of virtualized datacenters and cloud hosting services, the allocation and sizing of resources such as CPU, memory, and I/O bandwidth for virtual machines (VMs) is becoming increasingly important. Accurate performance modeling of an application would help users in better VM sizing, thus reducing costs. It can also benefit cloud service providers who can offer a new charging model based on the VMs' performance instead of their configured sizes. In this paper, we present techniques to model the performance of a VM-hosted application as a function of the resources allocated to the VM and the resource contention it experiences. To address this multi-dimensional modeling problem, we propose and refine the use of two machine learning techniques: artificial neural network (ANN) and support vector machine (SVM). We evaluate these modeling techniques using five virtualized applications from the RUBiS and Filebench suite of benchmarks and demonstrate that their median and 90th percentile prediction errors are within 4.36% and 29.17% respectively. These results are substantially better than regression based approaches as well as direct applications of machine learning techniques without our refinements. We also present a simple and effective approach to VM sizing and empirically demonstrate that it can deliver optimal results for 65% of the sizing problems that we studied and produces close-to-optimal sizes for the remaining 35%.

Miss rate curves (MRCs) are a fundamental concept in determining the impact of caches on an application's performance. In our research, we use MRCs to provision caches for applications in a consolidated environment. Current techniques for building MRCs at the CPU caches level require changes to the applications and are restricted to a few processor architectures [7], [22]. In this work, we investigate two techniques to partition shared L2 and L3 caches in a server and build MRCs for the VMs. These techniques make different trade-offs across accuracy, flexibility, and intrusiveness dimensions. The first technique is based on operating system (OS) page coloring and does not require change in commodity hardware or application. We improve upon existing page-coloring based approaches by identifying and overcoming a subtle but real problem of unequal associative cache sets loading to implement accurate cache allocation. Our second technique called Cache Grabber is even less intrusive and requires no changes in hardware, OS, or application. We present a comprehensive evaluation of the relative merits of these and other techniques to estimate MRCs. Our evaluation study enables a data center administrator to select the technique most suitable to his (her) specific data center to provision caches for consolidated applications.

Performance models provide the ability to predict application performance for a given set of hardware resources and are used for capacity planning and resource management. Traditional performance models assume the availability of dedicated hardware for the application. With growing application deployment on virtualized hardware, hardware resources are increasingly shared across multiple virtual machines. In this paper, we build performance models for applications in virtualized environments. We identify a key set of virtualization architecture independent parameters that influence application performance for a diverse and representative set of applications. We explore several conventional modeling techniques and evaluate their effectiveness in modeling application performance in a virtualized environment. We propose an iterative model training technique based on artificial neural networks which is found to be accurate across a range of applications. The proposed approach is implemented as a prototype in Xen-based virtual machine environments and evaluated for accuracy, sensitivity to the training process, and overhead. Median modeling error in the range 1.16-6.65% across a diverse application set and low modeling overhead suggest the suitability of our approach in production virtualized environments.

Duplication of data in storage systems is becoming increasingly common. We introduce I/O Deduplication, a storage optimization that utilizes content similarity for improving I/O performance by eliminating I/O operations and reducing the mechanical delays during I/O operations. I/O Deduplication consists of three main techniques: content-based caching, dynamic replica retrieval , and selective duplication . Each of these techniques is motivated by our observations with I/O workload traces obtained from actively-used production storage systems, all of which revealed surprisingly high levels of content similarity for both stored and accessed data. Evaluation of a prototype implementation using these workloads showed an overall improvement in disk I/O performance of 28 to 47% across these workloads. Further breakdown also showed that each of the three techniques contributed significantly to the overall performance improvement.

Applications that manage semi-structured data are becoming increasingly commonplace. Current approaches for storing semi-structured data use existing storage machinery; they either map the data to relational databases, or use a combination of flat files and indexes. While employing these existing storage mechanisms provides readily available solutions, there is a need to more closely examine their suitability to this class of data. Particularly, retrofitting existing solutions for semi-structured data can result in a mismatch between the tree structure of the data and the access characteristics of the underlying storage device (disk drive). This study explores various possibilities in the design space of native storage solutions for semi-structured data by exploring alternative approaches that match application data access characteristics to those of the underlying disk drive. For evaluating the effectiveness of the proposed native techniques in relation to the existing solution, we experiment with XML data using the XPathMark benchmark. Extensive evaluation reveals the strengths and weaknesses of the proposed native data layout techniques. While the existing solutions work really well for deep-focused queries into a semi-structured document (those that result in retrieving entire subtrees), the proposed native solutions substantially outperform for the non-deep-focused queries, which we demonstrate are at least as important as the deep-focused. We believe that native data layout techniques offer a unique direction for improving the performance of semi-structured data stores for a variety of important workloads. However, given that the proposed native techniques require circumventing current storage stack abstractions, further investigation is warranted before they can be applied to general-purpose storage systems.

Power consumption within the disk-based storage subsystem forms a substantial portion of the overall energy footprint in commodity systems. Researchers have proposed external caching on a persistent, low-power storage device, which we term external caching device (ECD), to minimize disk activity and conserve energy. While recent simulation-based studies have argued in favor of this approach, the lack of an actual system implementation has precluded answering several key questions about external caching systems. We present the design and implementation of EXCES, an external caching system that employs prefetching, caching, and buffering of disk data for reducing disk activity. EXCES addresses important questions related to external caching, including the estimation of future data popularity, I/O indirection, continuous reconfiguration of the ECD contents, and data consistency. We evaluated EXCES with both micro- and macro- benchmarks that address idle, I/O intensive, and real-world workloads. Overall system energy savings was found to lie in the modest 2–14% range, depending on the workload, in somewhat of a contrast to the higher values predicted by earlier studies. Furthermore, while the CPU and memory overheads of EXCES were well within acceptable limits, we found that flash-based external caching can substantially degrade I/O performance. We believe that external caching systems hold promise. Further improvements in ECD technology, both in terms of their power consumption and performance characteristics can help realize the full potential of such systems.

In a Grid computing environment, resources are shared among a large number of applications. Brokers and schedulers find matching resources and schedule the execution of the applications by monitoring dynamic resource availability and employing policies such as first-come-first-served and back-filling. To support applications with timeliness requirements in such an environment, brokering and scheduling algorithms must address an additional problem - they must be able to estimate the execution time of the application on the currently available resources. In this paper, we present a modeling approach to estimating the execution time of long-running scientific applications. The modeling approach we propose is generic; models can be constructed by merely observing the application execution “externally” without using intrusive techniques such as code inspection or instrumentation. The model is cross-platform; it enables prediction without the need for the application to be profiled first on the target hardware. To show the feasibility and effectiveness of this approach, we developed a resource usage model that estimates the execution time of a weather forecasting application in a multi-cluster Grid computing environment. We validated the model through extensive benchmarking and profiling experiments and observed prediction errors that were within 10% of the measured values. Based on our initial experience, we believe that our approach can be used to model the execution time of other time-sensitive scientific applications; thereby, enabling the development of more intelligent brokering and scheduling algorithms.

Database storage management at data centers is a manual, time-consuming, and error-prone task. Such management involves regular movement of database objects across storage nodes in an attempt to balance the I/O bandwidth utilization across disk drives. Achieving such balance is critical for avoiding I/O bottlenecks and thereby maximizing the utilization of the storage system. However, manual management of the aforesaid task, apart from increasing administrative costs, encumbers the greater risks of untimely and erroneous operations. We address the preceding concerns with STORM, an automated approach that combines low-overhead information gathering of database access and storage usage patterns with efficient analysis to generate accurate and timely hints for the administrator regarding data movement operations. STORM's primary objective is minimizing the volume of data movement required (to minimize potential down-time or reduction in performance) during the reconfiguration operation, with the secondary constraints of space and balanced I/O-bandwidth-utilization across the storage devices. We analyze and evaluate STORM theoretically, using a simulation framework, as well as experimentally. We show that the dynamic data layout reconfiguration problem is NP-hard and we present a heuristic that provides an approximate solution in O ( Nlog ( N / M ) + ( N / M ) 2 ) time, where M is the number of storage devices and N is the total number of database objects residing in the storage devices. A simulation study shows that the heuristic converges to an acceptable solution that is successful in balancing storage utilization with an accuracy that lies within 7% of the ideal solution. Finally, an experimental study demonstrates that the STORM approach can improve the overall performance of the TPC-C benchmark by as much as 22%, by reconfiguring an initial random, but evenly distributed, placement of database objects.

Database storage management in clustered storage environments is a manual, time-consuming, and error-prone task. Such management involves regular movement of database objects across nodes in the storage cluster so that storage utilization is maximized. We present STORM, an automated approach that guides this task by combining low-overhead information gathering about database access and storage usage patterns, efficient analysis of gathered information, and effective decision-making for reconfiguring data layout. The reconfiguration process is guided by the primary optimization objective of minimizing the total data movement required for the reconfiguration, with the secondary constraints of space and balanced I/O bandwidth utilizations across the storage nodes in the cluster. We model the reconfiguration decision-making as a multi-constraint optimization problem which is NP-hard. We then present a heuristic that provides an approximate solution in O(Nlog(N/M) + (N/M)2) time, where M is the number of storage nodes and N is the total number of database objects. A simulation study shows that the heuristic converges to an acceptable solution that is successful in balancing storage utilization with an accuracy that lies within 7% of the ideal solution.

Multimedia communication services today are conceived, designed, and developed in isolation, following a stovepipe approach. This has resulted in a fragmented and incompatible set of technologies and products. Building new communication services requires a lengthy and costly development cycle, which severely limits the pace of innovation. In this paper, we address the fundamental problem of automating the development of multimedia communication services. We propose a new paradigm for creating such services through declarative specification and generation, rather than through traditional design and development. Further, the proposed paradigm pays special attention to how the end-user specifies his/her communication needs, an important requirement largely ignored in existing approaches.

The performance of streaming media servers has been limited by the dual requirements of high disk throughput (to service more clients simultaneously) and low memory use (to decrease system cost). To achieve high disk throughput, disk drives must be accessed with large IOs to amortize disk access overhead. Large IOs imply an increased requirement of expensive DRAM, and, consequently, greater overall system cost. MEMS-based storage, an emerging storage technology, is predicted to offer a price-performance point between those of DRAM and disk drives. In this study, we propose storage architectures that use the relatively inexpensive MEMS-based storage devices as an intermediate layer (between DRAM and disk drives) for temporarily staging large disk IOs at a significantly lower cost. We present data layout mechanisms and synchronized IO scheduling algorithms for the real-time storage and retrieval of streaming data within such an augmented storage system. Analytical evaluation suggests that MEMS-augmented storage hierarchies can reduce the cost and improve the throughput of streaming servers significantly.

The development of collaborative multimedia applications today follows a vertical development approach, which is a major inhibitor that drives up the cost of development and slows down the pace of innovation of new generations of collaborative applications. In this paper, we propose a network communication broker (NCB) that provides a unified higher-level abstraction that encapsulates the complexity of network-level communication control and media delivery for the class of multimedia collaborative applications. NCB expedites the development of next-generation applications with diverse communication logics. Furthermore, NCB-based applications can be easily ported to new network environments. In addition, the self-managing design of NCB supports dynamic adaptation in response to changes in network conditions and user requirements

The convergence of data, voice and multimedia communication over digital networks, coupled with continuous improvement in network capacity and reliability has significantly enriched the ways we communicate. However, the stovepipe approach used to develop today's communication applications and tools results in rigid technology, limited utility, lengthy and costly development cycle, difficulty in integration, and hinders innovation. In this paper, we present a fundamentally different approach, which we call communication virtual machine (CVM) to address these problems. CVM provides a user-centric, model-driven approach for conceiving, synthesizing and delivering communication solutions across application domains. We argue that CVM represents a far more effective paradigm for engineering communication solutions. The concept, architecture, modeling language, prototypical design and implementation of CVM are discussed

Allowing higher-priority requests to preempt ongoing disk IOs is of particular benefit to delay-sensitive and real-time systems. In this paper, we present semi-preemptible IO, which divides disk IO requests into small temporal units of disk commands to improve the preemptibility of disk access. We first lay out main design strategies to allow preemption of each component of a disk access-seek, rotation, and data transfer, namely, seek-splitting, JIT-seek, and chunking. We then present the preemption mechanisms for single and multidisk systems-JIT-preemption and JIT-migration. The evaluation of our prototype system showed that semi-preemptible IO substantially improved the preemptibility of disk access with little loss in disk throughput and that preemptive disk scheduling could improve the response time for high-priority interactive requests.

The performance of streaming media servers has been limited due to the dual requirements of high throughput and low memory use. Although disk throughput has been enjoying a 40% annual increase, slower improvements in disk access times necessitate the use of large DRAM buffers to improve the overall streaming throughput. MEMS-based storage is an exciting new technology that promises to bridge the widening performance gap between DRAM and disk-drives in the memory hierarchy. We explore the impact of integrating these devices into the memory hierarchy on the class of streaming media applications. We evaluate the use of MEMS-based storage for buffering and caching streaming data. We also show how a bank of k MEMS devices can be managed in either configuration and that they can provide a k-fold improvement in both throughput and access latency. An extensive analytical study shows that using MEMS storage can reduce the buffering cost and improve the throughput of streaming servers significantly.

The use of interactive media has already gained considerable popularity. Interactivity gives viewers VCR controls like slow-motion, pause, fast-forward, and instant replay. However, traditional server-based or client-based approaches for supporting interactivity either consume too much network bandwidth or require large client buffering; and hence they are economically unattractive. We propose the architecture and design of an interactive media proxy (IMP) server that transforms noninteractive broadcast or multicast streams into interactive ones for servicing a large number of end users. For IMP to work cost-effectively, it must carefully manage its storage devices, which are needed for caching voluminous media data. In this regard, we propose a fine-grained device management strategy consisting of three complementary components: disk profiler, data placement, and IO scheduler. Through quantitative analysis and experiments, we show that these fine-grained strategies considerably improve device throughput under various workload scenarios.

This paper presents the architecture and implementation of XTREAM, a high-performance streaming multimedia system. XTREAM is supported by its three core components: IO scheduler, request handler, and admission controller. Via extensive experiments, we show that, thanks to these core components, XTREAM can achieve a low response time as well as high throughput and high-quality service to simultaneous clients.

Supporting preemptible disk access is essential for interactive multimedia applications that require short response time. In this study, we propose Virtual IO, an abstraction for disk IO, that transforms a non-preemptible IO request into a preemptible one. In order to achieve its objective efficiently, Virtual IO uses disk profiling to obtain accurate and detailed knowledge about the disk. Upon implementation of Virtual IO, we show that not only does Virtual IO enable highly preemptible disk access, but it does so with little or no loss in disk throughput.

In this paper, we propose an interactive DTV design that converts non-interactive broadcast DTV streams into interactive ones for multiple simultaneous viewers. To enable viewing interactivity, we show that it is critical to organize data intelligently for improving IO resolution, reducing disk latency and minimizing storage cost. We propose three data placement schemes that offer different tradeoffs between IO resolution, disk latency and storage. By employing different schemes under different workload scenarios, an intelligent system can minimize memory use and hence the system cost.