Angels: In Network Support for Minimum Distribution Time in P2P Overlays

This paper proposes the use of in-network caches (which we call Angels) to reduce the Minimum Distribution Time (MDT) of a file from a seeder – a node that possesses the file – to a set of leechers – nodes who are interested in downloading the file. An Angel is not a leecher in the sense that it is not interested in receiving the entire file, but rather it is interested in minimizing the MDT to all leechers, and as such uses its storage and up/down-link capacity to cache and forward parts of the file to other peers. We extend the analytical results by Kumar and Ross [1] to account for the presence of angels by deriving a new lower bound for the MDT. We show that this newly derived lower bound is tight by proposing a distribution strategy under assumptions of a fluid model. We present a GroupTree heuristic that addresses the impracticalities of the fluid model. We evaluate our designs through simulations that show that our Group-Tree heuristic outperforms other heuristics, that it scales well with the increase of the number of leechers, and that it closely approaches the optimal theoretical bounds.

Angels: In-Network Support for Minimum Distribution Time in P2P Overlays

This thesis proposes the use of in-network caches (which we call Angels) to reduce the Minimum Distribution Time (MDT) of a file from a seeder – a node that possesses the file – to a set of leechers – nodes who are interested in downloading the file. An Angel is not a leecher in the sense that it is not interested in receiving the entire file, but rather it is interested in minimizing the MDT to all leechers, and as such uses its storage and up/down-link capacity to cache and forward parts of the file to other peers. We extend the analytical results by Kumar and Ross (Kumar and Ross, 2006) to account for the presence of angels by deriving a new lower bound for the MDT. We show that this newly derived lower bound is tight by proposing a distribution strategy under assumptions of a fluid model. We present a GroupTree heuristic that addresses the impracticalities of the fluid model. We evaluate our designs through simulations that show that our GroupTree heuristic outperforms other heuristics, that it scales well with the increase of the number of leechers, and that it closely approaches the optimal theoretical bounds.

Concurrency Admission Control Management in ACCORD

We propose and evaluate admission control mechanisms for ACCORD, an Admission Control and Capacity Overload management Real-time Database framework-an architecture and a transaction model-for hard deadline RTDB systems. The system architecture consists of admission control and scheduling components which provide early notification of failure to submitted transactions that are deemed not valuable or incapable of completing on time. In particular, our Concurrency Admission Control Manager (CACM) ensures that transactions which are admitted do not overburden the system by requiring a level of concurrency that is not sustainable. The transaction model consists of two components: a primary task and a compensating task. The execution requirements of the primary task are not known a priori, whereas those of the compensating task are known a priori. Upon the submission of a transaction, the Admission Control Mechanisms are employed to decide whether to admit or reject that transaction. Once admitted, a transaction is guaranteed to finish executing before its deadline. A transaction is considered to have finished executing if exactly one of two things occur: Either its primary task is completed (successful commitment), or its compensating task is completed (safe termination). Committed transactions bring a profit to the system, whereas a terminated transaction brings no profit. The goal of the admission control and scheduling protocols (e.g., concurrency control, I/O scheduling, memory management) employed in the system is to maximize system profit. In that respect, we describe a number of concurrency admission control strategies and contrast (through simulations) their relative performance.

Factorization in the Cloud: Integer Factorization Using F# and Windows Azure

Implementations are presented of two common algorithms for integer factorization, Pollard’s “p – 1” method and the SQUFOF method. The algorithms are implemented in the F# language, a functional programming language developed by Microsoft and officially released for the first time in 2010. The algorithms are thoroughly tested on a set of large integers (up to 64 bits in size), running both on a physical machine and a Windows Azure machine instance. Analysis of the relative performance between the two environments indicates comparable performance when taking into account the difference in computing power. Further analysis reveals that the relative performance of the Azure implementation tends to improve as the magnitudes of the integers increase, indicating that such an approach may be suitable for larger, more complex factorization tasks. Finally, several questions are presented for future research, including the performance of F# and related languages for more efficient, parallelizable algorithms, and the relative cost and performance of factorization algorithms in various environments, including physical hardware and commercial cloud computing offerings from the various vendors in the industry.

Using Warp to Control Network Contention in Mermera

Parallel computing on a network of workstations can saturate the communication network, leading to excessive message delays and consequently poor application performance. We examine empirically the consequences of integrating a flow control protocol, called Warp control [Par93], into Mermera, a software shared memory system that supports parallel computing on distributed systems [HS93]. For an asynchronous iterative program that solves a system of linear equations, our measurements show that Warp succeeds in stabilizing the network's behavior even under high levels of contention. As a result, the application achieves a higher effective communication throughput, and a reduced completion time. In some cases, however, Warp control does not achieve the performance attainable by fixed size buffering when using a statically optimal buffer size. Our use of Warp to regulate the allocation of network bandwidth emphasizes the possibility for integrating it with the allocation of other resources, such as CPU cycles and disk bandwidth, so as to optimize overall system throughput, and enable fully-shared execution of parallel programs.

Dynamic Window-Constrained Scheduling for Real-Time Media Streaming

This paper describes an algorithm for scheduling packets in real-time multimedia data streams. Common to these classes of data streams are service constraints in terms of bandwidth and delay. However, it is typical for real-time multimedia streams to tolerate bounded delay variations and, in some cases, finite losses of packets. We have therefore developed a scheduling algorithm that assumes streams have window-constraints on groups of consecutive packet deadlines. A window-constraint defines the number of packet deadlines that can be missed in a window of deadlines for consecutive packets in a stream. Our algorithm, called Dynamic Window-Constrained Scheduling (DWCS), attempts to guarantee no more than x out of a window of y deadlines are missed for consecutive packets in real-time and multimedia streams. Using DWCS, the delay of service to real-time streams is bounded even when the scheduler is overloaded. Moreover, DWCS is capable of ensuring independent delay bounds on streams, while at the same time guaranteeing minimum bandwidth utilizations over tunable and finite windows of time. We show the conditions under which the total demand for link bandwidth by a set of real-time (i.e., window-constrained) streams can exceed 100% and still ensure all window-constraints are met. In fact, we show how it is possible to guarantee worst-case per-stream bandwidth and delay constraints while utilizing all available link capacity. Finally, we show how best-effort packets can be serviced with fast response time, in the presence of window-constrained traffic.

Task Assignment in a Distributed System: Improving Performance by Unbalancing Load

We consider the problem of task assignment in a distributed system (such as a distributed Web server) in which task sizes are drawn from a heavy-tailed distribution. Many task assignment algorithms are based on the heuristic that balancing the load at the server hosts will result in optimal performance. We show this conventional wisdom is less true when the task size distribution is heavy-tailed (as is the case for Web file sizes). We introduce a new task assignment policy, called Size Interval Task Assignment with Variable Load (SITA-V). SITA-V purposely operates the server hosts at different loads, and directs smaller tasks to the lighter-loaded hosts. The result is that SITA-V provably decreases the mean task slowdown by significant factors (up to 1000 or more) where the more heavy-tailed the workload, the greater the improvement factor. We evaluate the tradeoff between improvement in slowdown and increase in waiting time in a system using SITA-V, and show conditions under which SITA-V represents a particularly appealing policy. We conclude with a discussion of the use of SITA-V in a distributed Web server, and show that it is attractive because it has a simple implementation which requires no communication from the server hosts back to the task router.

Changes in Web Client Access Patterns: Characteristics and Caching Implications

Understanding the nature of the workloads and system demands created by users of the World Wide Web is crucial to properly designing and provisioning Web services. Previous measurements of Web client workloads have been shown to exhibit a number of characteristic features; however, it is not clear how those features may be changing with time. In this study we compare two measurements of Web client workloads separated in time by three years, both captured from the same computing facility at Boston University. The older dataset, obtained in 1995, is well-known in the research literature and has been the basis for a wide variety of studies. The newer dataset was captured in 1998 and is comparable in size to the older dataset. The new dataset has the drawback that the collection of users measured may no longer be representative of general Web users; however using it has the advantage that many comparisons can be drawn more clearly than would be possible using a new, different source of measurement. Our results fall into two categories. First we compare the statistical and distributional properties of Web requests across the two datasets. This serves to reinforce and deepen our understanding of the characteristic statistical properties of Web client requests. We find that the kinds of distributions that best describe document sizes have not changed between 1995 and 1998, although specific values of the distributional parameters are different. Second, we explore the question of how the observed differences in the properties of Web client requests, particularly the popularity and temporal locality properties, affect the potential for Web file caching in the network. We find that for the computing facility represented by our traces between 1995 and 1998, (1) the benefits of using size-based caching policies have diminished; and (2) the potential for caching requested files in the network has declined.