We analyze the design space for implementing the Constrained Application Protocol (CoAP) within Content-Centric Networks (CCN), identifying several CoAP specific scenarios and seeing how they map to CCNs. We present an evaluation, recommendations for implementations and extensions, and directions for future work. Our key result is that while several protocol features and flows map naturally, some communication patterns are more difficult to capture without modifications or additions to the CCN model. In addition to many conceptual and also non-performance related advantages, our experimental evaluation demonstrates that it is feasible and useful to implement CoAP over CCN, with the performance similar to that of CoAP over UDP over IP, and in several cases outperforming it significantly.

An Internet of Things (IoT) environment typically consists of lossy and low powered networks (LLN) of interconnected sensors. An IoT messaging protocol that is gaining in popularity and importance is the Constrained Application Protocol (CoAP). CoAP runs at the application layer, handling reliability and congestion control over the unreliable UDP layer. Due to low bandwidth and high scale of communication, congestion can occur among sensors in an LLN, or when communicating to a border router, or when clients from the Internet access the resources in an LLN. While CoAP was originally specified with a primitive congestion control mechanism, to address its various limitations a new congestion control protocol named CoAP Congestion Control Advanced (CoCoA) is being developed. We present an analysis and evaluation of CoCoA, and based on our findings we propose and implement CoCoA 4-state-Strong, an adaptation of CoCoA that uses a 4-state estimator for variable backoffs to significantly improve throughput, even in highly lossy networks, while maintaining a higher goodput.

We describe a new enhanced cloud computing architecture, called the Proximal Workspace, to allow access and interaction between lightweight wearable I/O devices, e.g., video glasses, earphones, wrist displays, body sensors, etc., and ubiquitous applications that represent a new generation of comp

With applications increasingly moving to the cloud, it is becoming common for an application to be separated by the network from the I/O devices with which the user is interacting. Currently this requires modifying the application to receive user input from the network rather than the device. We present a new I/O architecture in which the device driver is split into two parts, with the network between them. This architecture makes the network invisible to both device and application, allowing both of them to work unmodified. Our architecture also supports transformation modules, each of which comes in a pair that operates on each side of the network. Via these module pairs, the resulting system is capable of supporting the modification of the I/O stream in a variety of ways to compensate for the network, while remaining transparent to the application.

Predictive methods for learning network distances are often more desirable than direct performance measurements between end hosts. Yet, predicting network distances remains an open and difficult problem, as the results from a number of comparative and analytical studies have shown. From an application requirements perspective, there is significant room for improvement in achieving prediction accuracies at a satisfactory level. In this paper, we develop and analyze a new, machine learning-based approach to distance prediction that seeks to capture and generalize geographical characteristics between Internet node pairs, instead of relying on direct and ongoing measurements of partial paths. We apply a basic algorithm in machine learning to demonstrate this idea and highlight the potential benefits that this method may offer over other popular methods that exist today.

We present performance results of a new distributed I/O software architecture to support remote applications interacting with local I/O devices. The architecture emphasizes network transparency and ease of customization/extensibility in support of the vastly different needs of various applications and devices that can benefit from remote I/O. Networked I/O is achieved via a networked device driver that is split into two parts, one on each side of the network. An I/O stream that is sourced at one end and sinked at the other may be modified by a set of pipelined transformation modules. Each module comes in a pair, one on each side of the network, with one side typically applying some operation and the other side applying a corresponding one, such as encoding and decoding the format of the data or pausing and resuming the sending of messages. Because of the paired nature of transformation modules, the system is capable of supporting the modification of the I/O stream in a variety of ways to compensate for network issues, one of the key problems of remote I/O, while remaining transparent to the application. We show that even with an implementation that operates almost entirely at user level (i.e., outside the operating system), good levels of performance that are adequate for even high intensity I/O, both in terms of efficiency and throughput, can be achieved.

Replicated storage systems allow their stored data objects to outlive the life of the nodes storing them through replication. In this paper, we focus on durability, and more specifically on the concept of an object's lifetime, i.e., the duration of time between the creation of an object and when it is permanently irretrievable from the system. We analyze two main replication strategies: reactive, in which replication occurs in response to failures, and proactive, in which replication occurs in anticipation of failures. Our work presents a quantitative analysis that compares reactive and proactive through analytical models and simulations, considering exponentially distributed failures and reactive repairs, and periodic proactive replications. We also present a derivation of the analytical formula for the variance of the lifetime in the reactive model. Our results indicate that a proactive strategy leads to multiple times higher storage requirements than a reactive strategy. In addition, reactive systems are only moderately bursty in terms of bandwidth consumption, with rare peaks of at most five times the bandwidth consumption in proactive systems (given input parameter values that are compatible with real systems). Finally, for both strategies, the standard deviation is very close to the expected lifetime, and consequently, the lifetimes close to being exponentially distributed.

In this paper, we study nodes from the PlanetLab test bed to form a model of their uptime behavior. By applying clustering techniques to over a year's worth of availability data for the nodes, we identify six uptime distributions, each exhibiting unique characteristics shared by the nodes within it. The behavioral patterns exhibited by these groups, combined with the behaviors exhibited by the aggregate across the system, provide useful information for researchers designing applications that are run or tested on PlanetLab.

Tor (The Onion Router) is a popular anonymity overlay network that seeks to provide anonymity without significant cost to performance. Tor's support for anonymity is indeed strong, but its network performance is a problem, and one that is widely recognized. While there are some studies that investigate changing the structure of the Tor network to improve performance, we focus on investigating different path selection strategies given the Tor network as is. Specifically, we explore varying the number of hops in a circuit, varying the performance flags in a circuit, and varying the geographic distance between routers in a circuit. We show how much improvement can be had by reducing the path length, which gives the user guidance on how to trade off anonymity for performance.

VNC (Virtual Network Computing), a form of thin client computing, offers many advantages over traditional desktop computing including lower costs, higher security, ubiquitous access to resources, and easier maintenance. In this paper, we focus on the problems of using VNC to display video over a high-latency network, and how to solve them. VNC performance is inherently capped at 1 frame per round-trip time (between the client and server) due to its client-pull communication style. We describe VNC-HL, our extension of VNC, which improves frame rate performance by employing a pre-requesting technique to periodically request updates, even with several previous requests pending. Experimental results demonstrate up to an order of magnitude of improvement. VNC-HL achieved 14 FPS over a network with 500 ms of latency.

In this paper, we investigate the roles of replication vs. repair to achieve durability in large-scale distributed storage systems. Specifically, we address the fundamental questions: How does the lifetime of an object depend on the degree of replication and rate of repair, and how is lifetime maximized when there is a constraint on resources? In addition, in real systems, when a node becomes unavailable, there is uncertainty whether this is temporary or permanent; we analyze the use of timeouts as a mechanism to make this determination. Finally, we explore the importance of memory in repair mechanisms, and show that under certain cost conditions, memoryless systems, which are inherently less complex, perform just as well.

We present network flow buffering, which is the use of a remote flow-regulating buffer that is deployed between a Web client and server to improve performance of HTTP-based playback of video. We show that HTTP enhanced with network flow buffering significantly improves performance, especially under high packet loss and highly variable bandwidth conditions, when compared with using either straight HTTP or streaming. The benefits of network flow buffering are numerous, and include reducing interruptions to playback, maintaining high video-image quality, and decreasing client-buffering requirements. Network flow buffering is easily implemented with existing Web mechanisms.

We study the problem of guaranteeing data durability [2] in distributed storage systems based on replication. Our work is motivated by several several recent efforts [3, 5, 1] to build such systems in a peer-to-peer environment. The key features of this environment which make achieving durability difficult are (1) data lifetimes may be several orders of magnitude larger than the lifetimes of individual storage units, and (2) the system may have little or no control over the participation of these storage units in the system. We use a model-based approach to develop engineering principles for designing automated replication and repair mechanisms to implement durability in such systems.

Indirect routing involves sending messages between Internet end nodes through a specified intermediate node to effect a different end-to-end route than the default "direct" route. Indirect routing offers the potential for significant improvements in throughput performance by exploiting the Internet's richness in throughput diversity. We present a performance analysis of indirect routing, showing that it can result in a 33-49% increase in average throughput performance while incurring low overhead.

This paper focuses on peer-to-peer storage systems that achieve availability through replication. We study the problem of resource allocation when the system must replicate multiple files using a fixed amount of resource. We characterize the optimal allocation that maximizes the average availability of the files in the system, and also study two simple, decentralized allocation schemes, viz., uniform allocation, where each file is allocated equal shares of the resource, and proportional allocation, where each file is allocated a share of the resource proportional to its size. We show that while uniform allocation is fair in terms of allocating resources, it may be arbitrarily sub-optimal. On the other hand, proportional allocation, though unfair in resource allocation, is competitive with the optimal allocation.

ALPS is a per-application user-level proportional-share scheduler that operates with tow overhead and without any special kernel support. ALPS is useful to a range of applications, including scientific applications that need to control the CPU apportionment to the processes they create, to Web servers that need to limit the proportion of available CPU time given to spawned processes that service Web requests, and to middleware that supports multiple execution environments that are to run at different rates. ALPS works by minimally sampling the progress of processes under its control, and making simple predictions for when it should selectively pause and resume the processes. We present the algorithm, a UNIX-based implementation, and a performance evaluation. Our results show that the ALPS approach is practical; we can achieve good accuracy (under 5% error), and low overhead (under 1% of CPU), despite user-level operation

Abstract — We address the problem of using replication to reliably maintain state in a distributed system for time spans that far exceed the lifetimes of individual replicas. This scenario is relevant for any system comprised of a potentially large and selectable number of replicated components, each of which may be highly unreliable, where the goal is to have enough replicas to keep the system “alive ” (meaning at least one replica is working or available) for a certain expected period of time, i.e., the system’s lifetime. In particular, this applies to recent efforts to build highly available storage systems based on the peer-to-peer paradigm. We model notions of replica loss and replica repair in such systems by a simple Markov chain model, and derive an expression for the lifetime of the replicated state. We then apply this model to study the impact of practical considerations like storage and bandwidth limits on the system, and describe methods to optimally choose system parameters so as to maximize lifetime. Our analysis sheds light on the efficacy of various replication strategies. I.

Many overlay multicast protocols have been designed and deployed across the Internet to support content distribution. To our knowledge, however, none have provided a rigorous analysis of the problem or the effectiveness of their proposed solutions. We define the Overlay Network Content Distribution (OCD) problem to allow such analyses.

ReAgents are remotely executing agents that customize Web browsing for non-standard clients. A reAgent is essentially a one-shot" mobile agent that acts as an extension of a client dynamically launched by the client to run on its behalf at a remote more advantageous location. ReAgents simplify the use of mobile agent technology by transparently handling data migration and run-time network communications and provide a general interface for programmers to more easily implement their application-specific customizing logic. This is made possible by the identification of useful remote behaviors i.e. common patterns of actions that exploit the ability to process and communicate remotely. Examples of such behaviors are transformers monitors cachers and collators. In this paper we identify a set ofuseful reAgent behaviors for interacting with Web services via astandard browser describe how to program and use reAgents and show that the overhead of using reAgents is low and outweighed by its benefits.

Fair queuing is a well-studied problem in modern computer networks. However, there remains a gap between scheduling algorithms that have provably good performance, and those that are feasible and practical to implement in high speed routers. In this paper, we propose a novel packet scheduler called Stratified Round Robin, which has low complexity, and is amenable to a simple hardware implementation. Stratified Robin Robin exhibits good fairness and delay properties that are demonstrated through both analytical results and simulations. In particular, it provides a single packet delay bound that is independent of the number of flows. This property is unique to Stratified Round Robin among all other schedulers of comparable complexity.

We present a new Web middleware architecture that allows users to customize their view of the Web for optimal interaction and system operation when using non-traditional resource-limited client machines such as wireless PDAs (personal digital assistants). Web Stream Customizers (WSC) are dynamically deployable software modules and can be strategically located between client and server to achieve improvements in performance, reliability, or security. An important design feature is that Customizers provide two points of control in the communication path between client and server, supporting adaptive system-based and content-based customization. Our architecture exploits HTTP's proxy capabilities, allowing Customizers to be seamlessly integrated with the basic Web transaction model. We describe the WSC architecture and implementation, and illustrate its use with three non-trivial, adaptive Customizer applications that we have built. We show that the overhead in our implementation is small and tolerable, and is outweighed by the benefits that Customizers provide.

Providing quality of service (QoS) guarantees for wireless communications requires special consideration for the relatively large and time-varying bit error rates of wireless links, due to impairments that are difficult to predict. We describe a new wireless link access scheme called remote-queueing multiple access (RQMA), which allows seamless integration of scheduling disciplines proposed for wired networks that achieve end-to-end QoS guarantees. RQMA supports the flexible and efficient allocation of bandwidth for real-time sessions. Mobile stations send information about their backlogged packets to a base station which executes a scheduling discipline that provides delay bounds. Some minimum portion of the link bandwidth is reserved for the retransmission of packets that are not correctly received but whose deadlines are still not violated. For a fast fading channel with an average error rate of 10/sup -2/, RQMA is capable of delivering real-time traffic with a small average packet loss rate on the order of 10/sup -5/. Similar performance can be achieved for a slow fading channel if the average error rate is no more than 10/sup -3/.

Rate-function scheduling (RFS) is a new deadline-based packet-scheduling service discipline that supports quality-of-service guarantees for applications with real-time communication requirements. RFS is distinguished from other service disciplines in that it achieves all of the following goals: analytically-derived performance bounds, performance isolation among sessions, flexible and efficient allocation of bandwidth, implementation simplicity, work-conserving operation, and bandwidth-fair operation (defined in the paper). Through the specification of rate functions, sessions can control their bandwidth usage and their upper bounds on delay. For a class of rate functions, which we show is sufficient for providing sessions with delay bounds, RFS is as simple to implement and to calculate service bounds such as Zhang's VirtualClock service discipline. We also show that the non-preemptive earliest deadline first policy is a simple degeneration of RFS.

In a deadline-ordered service discipline, packets are assigned transmission deadlines and eligibility times and are transmitted in increasing order of deadlines. Different deadline-ordered service disciplines are distinguished by how they calculate deadlines and eligibility times. One of the more difficult analytical problems one faces when designing a new deadline-ordered service discipline is to prove that one can bound the end of transmission times of packets relative to their assigned deadlines, which we call schedulability. We show that, no matter how one calculates deadlines, there is a simple schedulability condition for deadline-ordered service disciplines. This schedulability condition is necessary and sufficient for preemptive deadline-ordered service disciplines, and for a server that allows the presence of nonreal-time packets (i.e., packets with no deadlines), it is also necessary and sufficient for nonpreemptive deadline-ordered service disciplines. We also address the schedulability problem for service disciplines in general, and show the optimality of deadline-ordered service disciplines. To demonstrate how our results simplify schedulability determination, we use them to prove the known schedulability conditions of VirtualClock, Packet-by-Packet Generalized Processor Sharing (PGPS), Stop-and-Go, and Delay-Earliest-Due-Date (Delay-EDD), and to provide a new result, the necessary schedulability condition of VirtualClock.

This paper presents detailed measurements of processing overheads for the Ultrix 4.2a implementation of TCP/IP network software running on a DECstation 5000/200. The performance results were used to uncover throughput and latency bottlenecks. We present a scheme for improving throughput when sending large messages by avoiding most checksum computations in a relatively safe manner. We also show that for the implementation we studied, reducing latency (when sending small messages) is a more difficult problem because processing overheads are spread over many operations; gaining a significant savings would require the optimization of many different mechanisms. This is especially important because, when processing a realistic workload, we have found that nondata-touching operations consume more time in aggregate than data-touching operations.

In autonomous decentralized systems (ADS), there is a high degree of uncertainty, especially in global state information and the performance and reliability of various resources. We model a distributed computer system of local and shared global servers, each of which can process jobs. The local servers insure a degree of autonomy and the global servers provide added computational power and redundancy. When a job arrives, an agent decides whether to schedule the job locally, or whether to ship it to a global server. The agents have access to state information concerning current loads on the global servers and the response times of completed jobs, but because of decentralization this information is delayed, and perhaps lost, along communication channels. The agents must make good decisions under the constraint that service rates are unknown and dynamic. We evaluate a deterministic algorithm and two randomizing algorithms and show that the degree of uncertainty determines which one is the most successful.>

Leave-in-Time is a new rate-based service discipline for packet-switching nodes in a connection-oriented data network. Leave-in-Time provides sessions with upper bounds on end-to-end delay, delay jitter, buffer space requirements, and an upper bound on the probability distribution of end-to-end delays. A Leave-in-Time session's guarantees are completely determined by the dynamic traffic behavior of that session, without influence from other sessions. This results in the desirable property that these guarantees are expressed as functions derivable simply from a single fixed-rate server (with rate equal to the session's reserved rate) serving only that session. Leave-in-Time has a non-work-conserving mode of operation for sessions desiring low end-to-end delay jitter. Finally, Leave-in-Time supports the notion of delay shifting, whereby the delay bounds of some sessions may be decreased at the expense of increasing those of other sessions. We present a set of admission control algorithms which support the ability to do delay shifting in a systematic way.

No abstract available.

Proves that a connection composed of virtual-clock servers provides an upper bound on delay for leaky bucket constrained sessions, i.e., sessions conforming to a token bucket filter. This upper bound on delay is calculated, and it is the same upper bound on delay given by PGPS. The authors also prove that leaky bucket constrained sessions are the only type of sessions for which an upper bound on delay can be provided by servers with an upper bound on link capacity.>

In a distributed system, a group of agents have a potential for improved performance depending on their ability to utilize shared resources. This potential synergy raises the question of whether agents should work together in a system-wide group, i.e., a coalition, or whether they should work alone. In general, there is uncertainty as to whether a coalition will form; this uncertainty can arise for various reasons, such as adaptive strategies of the agents or random faults in the system. In this paper, we present a model for performance based upon the probability of coalition formation. The results indicate a limit in potential performance for adaptive agents and, in particular, the global and local maxima along with regions of nonstability. In addition, the model shows how performance is affected by the knowledge environment of the distributed system, that is, the architecture of the system with respect to the distribution of information. Four environments are examined as illustrations of these general categories of information distribution: global information; inaccessible information; local information residing in autonomous agents; and information residing in a master control agent. The results show the distinctions between the environments with respect to probabilistic coalition formation and also demonstrate the loss in environments without communication as compared to a baseline communication environment.>

A new model for distributed decision making, distributed game automata, focuses on how communication affects the quality of decisions. The goal is to limit communication between decision makers such that overhead costs are reduced but good decisions still result. Learning automata play repeated games with payoffs quantifying the performance in a distributed application. Each automaton's view of the global state is periodically updated by communication from other automata of their local state (i.e. current strategy probabilities). Because of infrequent communication and transmission delays, received state information may become stale: an example game illustrates the mutually conflicting decisions that can result. Simulation and analytic results show there exists a maximum communication delay before decision quality begins to suffer, however, with sufficient communication, the agents adapt to a coordinated policy.>

Each job scheduling agent in large decentralized load balancing systems generally has a set of remote hosts (i.e., its scope of control) to consider for offloading when the local load is too high. Typically, each agent's scope of control includes all the hosts in the system. The potential performance benefits of limiting the size of each agent's scope of control are investigated. The larger the scope of control, the less often state information can lower the quality of load balancing decisions. The smaller the scope of control, the less opportunity an agent has for finding a lightly loaded host for offloading. Agents adaptively modify their scope of control over time based on feedback regarding the success of load balancing decisions and act as a self-organizing system to efficiently share available processing power.>

The performance of a randomized coordination mechanism for resource access in an autonomous decentralized system is investigated. Given n users and r resources, each user decides to access a resource with probability alpha , and each resource is selected with equal likelihood. The goal is to select alpha so that mutually conflicting decisions, where multiple users try to access the same resource, are minimized. Results are presented on the optimal number of users given some number of resources, and the limiting performance as the number of users approaches infinity. It is shown that given n = kr users and r resources, it is better to partition users and resources so that each resource is accessed by a separate group of k users rather than to allow all n users to compete for all r resources.>

We present detailed measurements of various processing overheads of the TCP/IP and UDP/IP protocol stacks on a DECstation 5000/200 running the Ultrix 4.2a operating system. These overheads include data-touching operations, such as the checksum computation and data movement, which are well known to be major time consumers. In this study, we also considered overheads due to non-data touching operations, such as network buffer manipulation, protocol-specific processing, operating system functions, data structure manipulations (other than network buffers), and error checking. We show that when one considers realistic message size distributions, where the majority of messages are small, the cumulative time consumed by the non-data touching overheads represents the majority of processing time. We assert that it will be difficult to significantly reduce the cumulative processing time due to non-data touching overheads.

Network software speed is not increasing as rapidly as that of work-station CPUs. The goal of this study is to determine how various components of network software contribute to this bottleneck. In this extended abstract, we summarize the performance of TCPIP and UDPIP networking software for the DECstation 5000/200 workstations connected by an FDDI LAN.

The authors present heuristics for multicast tree construction for communication that depends on: bounded end-to-end delay along the paths from source to each destination and minimum cost of the multicast tree, where edge cost and edge delay can be independent metrics. The problem of computing such a constrained multicast tree is NP-complete. It is shown that the heuristics demonstrate good average case behavior in terms of cost, as determined by simulations on a large number of graphs.>

We investigate how delayed communication affects the dynamic formation of groups in distributed systems, where all decision-making agents join the same group because each expects to improve its own performance. For example, distributed job schedulers may form a group to utilize the idle resources of other members within the group. Forming a group is a search problem and we examine agents that use the feedback mechanism of stochastic learning automata to carry out this search. Although a group formation may have the potential for synergy, the agents must successfully coordinate their actions within the group relevant to the application. For example, job schedulers who form a group must still balance the load among the shared resources; that is, the collective actions of the schedulers need to be coordinated and greedy schedulers who all pick the same processor may not be successful. Agents may find that working alone is more desirable since their actions need not be coordinated and the results of their own actions are more predictable. An additional challenge to the search problem is to cope with the delay in communication between the agents. The purpose of this study is to model systems where agents adaptively search for compatible co-workers, under the constraint of delayed communication. With insufficient communication, the agents decide to work alone (and receive a modest benefit) but, with sufficient communication, the agents make the more advantageous decision to work together.>

A simple system that demonstrates coadaptation in distributed decision-making as applied to the problem of job scheduling is presented. The decision-making process is constructed as a set of autonomous agents based on learning automata that can individually learn from feedback. The set of agents adapt together, or coadapt, to form a good, though not necessarily optimal, global decision process. Coadaptive distributed computer job scheduling in a two-machine system is simulated with the agents implemented as stochastic learning automata. It is demonstrated that the performance of these coadaptive agents is similar to that of a related group of well-performing static decision-making agents, that quantitative changes in agents' parameters cause qualitative changes in the coadaptive behavior, and that constructing an optimal agent for a coadapting system is dependent upon the other agents present. Coadaptive behavior is clearly affected by the relative frequencies with which decisions are made, and relative sizes of the rewards and penalties.>

The authors investigate multicast routing for high-bandwidth delay-sensitive applications in a point-to-point network as an optimization problem. They associate an edge cost and an edge delay with each edge in the network. The problem is to construct a tree spanning the destination nodes, such that it has the least cost, and so that the delay on the path from the source to each destination is bounded. Since the problem is computationally intractable, the authors present an efficient approximation algorithm. Experimental results through simulations show that the performance of the heuristic is near optimal.>

The San Diego Supercomputer Center is one of four NSF sponsored national supercomputer centers. Up until January of 1990, its workhorse was a Cray X-MP, which served 2700 researchers from 170 institutions, spanning 44 states. In order to better understand how this supercomputer was utilized by its diverse community of users, we undertook a workload characterization study of the Cray X-MP. The goals of our study were twofold. First, we wished to characterize the workload at both the functional and resource levels. The functional level represents the user point of view: what types of programs users are running on the system. The resource level represents the system point of view: how the systems resources (CPU, memory, I/O bandwidth) are being used. Second, we wanted to see how the workload changed over an average weekday. Thus, we conducted a static characterization to understand its global attributes over the entire measurement period, as well as a dynamic workload characterization to understand the time behavior of the workload over a weekday cycle.

File system I/O is increasingly becoming a performance bottleneck in large distributed computer systems. This is due to the increased file I/O demands of new applications, the inability of any single storage structure to respond to these demands, and the slow decline of, disk access times (latency and seek) relative to the rapid increase in CPU speeds, memory size, and network bandwidth.We present a multi-structured file system designed for high bandwidth I/O and fast response. Our design is based on combining disk caching with three different file storage structures, each implemented on an independent and isolated disk array. Each storage structure is designed to optimize a different set of file system access characteristics such as cache writes, directory searches, file attribute requests or large sequential reads/writes.As part of our study, we analyze the performance of an existing file system using trace data from UNIX disk I/O-intensive workloads. Using trace driven simulations, we show how performance is improved by using separate storage structures as implemented by a multi-structured file system.