Transmission schemes that gain content from multiple servers concurrently have been highlighted due to their ability to provide bandwidth aggregation, stability on dynamic server departure, and load balancing. Previous approaches employ parallel downloading in the transport layer to minimize the receiver buffer size and maximize bandwidth utilization. However, they only focus on the receiver operations and induce considerable overhead at the senders in contradiction to the main goal of a multi-provider environment, offloading popular servers through replication. In the present work, the authors propose MTCP, a novel transport layer protocol that focuses on reduction of the sender overhead through the elimination of unnecessary disk I/Os and efficient buffer cache utilization. MTCP also balances trade-off objectives to minimize buffering at receivers and maximize the request locality at senders.

Many cooperated web cache systems and protocols have been proposed. These systems, however, require expensive resources, such as external bandwidth and CPU power or storage of a proxy, while inducing hefty administrative costs to achieve adequate client population growth. Moreover, a scalability problem in the cache server management still exists. This paper suggests peer-to-peer client-clustering. The client-cluster provides a proxy cache with backup storage which is comprised of the residual resources of the clients. We use DHT based peer-to-peer lookup protocol to manage the client-cluster. With the natural characteristics of this protocol, the client-cluster is self-organizing, fault-tolerant, well-balanced and scalable. Additionally, we propose the Backward ICP which is used to communicate between the proxy cache and the client-cluster, to reduce the overhead of the object replication and to use the resources more efficiently. We examine the performance of the client-cluster via a trace driven simulation and demonstrate effective enhancement of the proxy cache performance.

DHT based p2p systems appear to provide scalable storage services with idle resource from many unreliable clients. If a DHT is used in storage intensive applications where data loss must be minimized, quick replication is especially important to replace lost redundancy on other nodes in reaction to failures. To achieve this easily, a simple replication method directly uses a consistent set, such as a leaf set and a successor list. However, this set is tightly coupled to the current state of nodes and the traffic needed to support this replication can be high and bursty under churn. This paper explores efficient replication methods that only glimpse a consistent set to select a new replica. Replicas are loosely coupled to a consistent set and we can eliminate the compulsory replication under churn. Because of a complication of the new replication methods, the careful data management is needed under churn for the correct and efficient data lookup. Results from a simulation study suggest that our methods can reduce network traffic enormously for high data durability.

While web proxy caching is a widely deployed technique, the performance of a proxy cache is limited by the local storage. Some studies have addressed this limitation by using the residual resources of clients via a p2p method and have achieved a very high hit rate. However, these approaches treat web objects as homogeneous objects and there is no consideration of various web characteristics. Consequently, the byte hit rate of the system is limited, external bandwidth is wasted, and perceived user latency is increased. The present paper suggests an efficient p2p based web caching technique that manages objects with different policies so as to exploit the characteristics of web objects, such as size and temporal locality. Small objects are stored alone whereas large objects are stored by dividing them into numerous small blocks, which are distributed in clients. On a proxy cache, header blocks of large objects take the place of objects themselves and smaller objects are cached. This technique increases the hit rate. Unlike a web cache, which evicts large objects as soon as possible in the case where clients fulfill the role of backup storage, large objects are given higher priority than small objects in the proposed approach. This maximizes the effect of hits for large objects and thereby increases the byte hit rate. Furthermore, we construct simple latency models for various p2p based web caching systems and analyze the effects of the proposed policies on these systems. We then examine the performances of the efficient policies via a trace driven simulation. The results demonstrate that the proposed techniques effectively enhance web cache performance, including hit rate, byte hit rate, and response time.

With advances in ubiquitous environments, user demand for easy data-lookup is growing rapidly. Not only users but intelligent ubiquitous applications also require data-lookup services for a ubiquitous computing framework. This paper proposes a backward-compatible, searchable virtual file system (S-VFS) for easy data-lookup. We add search functionality to the VFS, the de facto standard abstraction layer over the file system. Users can find a file by its attributes without remembering the full path. S-VFS maintains the attributes and the indexing structures in a normal file per partition. It processes queries and returns the results in a form of a virtual directory. S-VFS is the modified VFS, but uses legacy file systems without any modification. Since S-VFS supports full backward compatibility, users can even browse hierarchically with the legacy path name. We implement S-VFS in Linux kernel 2.6.7-21. Experiments with randomly generated queries demonstrate outstanding lookup performance with a small overhead for indexing.

Shared memory multiprocessors in which each processor has its own TLB must manage consistency among TLBs and a page table. As the large-scale CC-NUMA (cache-coherent non-uniform memory access) shared memory multiprocessors become popular, it is important for TLB consistency management algorithms to be highly scalable. In this paper, we propose a TLB update-hint algorithm as a scalable TLB consistency management solution for CC-NUMA multiprocessors. By using a lazy TLB invalidation approach, we reduced the number of unnecessary processor interruptions and idle-waiting time, and achieved a high level of scalability. Using a shared memory simulator, we evaluated the TLB update-hint algorithm. For performance comparison, we also simulated the TLB shootdown algorithm, one of the most popular TLB consistency algorithms. The simulations demonstrated that the TLB update-hint algorithm scales well in systems with a large number of processors. At 64 node systems, the TLB update-hint algorithm shows 4787% better performance than the TLB shootdown algorithm.

Global computing system (GCS) harnesses the idle CPU resources of clients connected to Internet for solving large problems that require high volume of computing power. Since GCS scale to millions of clients, many projects usually adopt coarse-grained scheduling in order to reduce server-side contention at the expense of sacrificing the degree of parallelism and wasting CPU resources. In this paper, we propose a new type of client, i.e., a scheduling proxy that enables adaptive-grained scheduling between the server and clients. While the server allocates coarse-grained work units to scheduling proxies alone, clients download fine-grained work units from a relatively nearby scheduling proxy not from the distant server. By computation of small work units at client side, the turnaround time of work unit can be reduced and the waste of CPU time by timeout can be minimized without increasing the performance cost of contention at the server. In addition, in order not to lose results in the failure of scheduling proxies, we suggest a technique of result caching in clients.

While superpages are an efficient solution to increase TLB reach, strong contraint for using superpages hinders the actual utilization. Two previous solutions, a partial-subblock TLB and the shadow memory were proposed to loose the contraint. A partial-subblock TLB looses only a small portion of the contraint and limits the superpage size at a cost. The shadow memory looses most of the constraint but introduces other serious problems. We propose three novel approaches to improve superpage supports. First, we propose a hybrid scheme which integrates both the shadow memory and a partial-subblock TLB, thereby enjoying the benefits inherited from both sides. The hybrid scheme has as high a superpage utilization as the shadow memory, and avoids most of the problems in the shadow memory by the virtue of partial-subblock TLB. Second, VS-TLBs are an extension of subblock TLBs to support multiple page subblocks, while subblock TLBs can support only single page subblocks. VS-TLBs have a much larger TLB reach than subblock TLBs with a cost of a small number of bits. Last, we propose VS-hybrid which replaces the partial-subblock TLB in the hybrid scheme with a partial VS-TLB. It supports multiple page subblocks in the hybrid scheme. Therefore, it takes both advantages of the hybrid scheme and the expanded subblock size. The simulation results show that the proposed schemes take a large amount of performance gain in the benchmark application programs.

Content-based publish/subscribe systems provide a useful alternative to traditional address-based communication due to their ability to decouple communication between participants. It has remained a challenge to design a scalable overlay supporting the complexity of content-based networks, while satisfying the desirable properties large distributed systems should have. This paper presents the design of Mirinae, a new structured peer-to-peer overlay mesh based on the interests of peers. Given an event, Mirinae provides a flexible and efficient dissemination tree minimizing the participation of non-matching nodes. We also present a novel ID space transformation mechanism for balancing routing load of peers even with highly skewed data, which is typical of the real world. Our evaluation demonstrates that Mirinae is able to achieve its goals of scalability, efficiency, and near-uniform load balancing. Mirinae can be used as a substrate for content-search and range query in other important distributed applications.

In traditional file systems, data clustering and grouping have improved small file performance. These schemes make it possible for file systems to use large data transfers in accessing small files, reducing disk I/Os. However, as file systems age, disks become too fragmented to support the grouping and clustering. To offer a solution to this problem, we describe a De-Fragmented File System (DFFS), which gradually alleviates fragmentation of small files. By using data cached in memory, DFFS dynamically relocates blocks of small fragmented files, clustering them on the disks contiguously. In addition, DFFS relocates small related files in the same directory, grouping them at contiguous disk locations.