Mermera: Non-Coherent Distributed Shared Memory for Parallel Computing

The proliferation of inexpensive workstations and networks has prompted several researchers to use such distributed systems for parallel computing. Attempts have been made to offer a shared-memory programming model on such distributed memory computers. Most systems provide a shared-memory that is coherent in that all processes that use it agree on the order of all memory events. This dissertation explores the possibility of a significant improvement in the performance of some applications when they use non-coherent memory. First, a new formal model to describe existing non-coherent memories is developed. I use this model to prove that certain problems can be solved using asynchronous iterative algorithms on shared-memory in which the coherence constraints are substantially relaxed. In the course of the development of the model I discovered a new type of non-coherent behavior called Local Consistency. Second, a programming model, Mermera, is proposed. It provides programmers with a choice of hierarchically related non-coherent behaviors along with one coherent behavior. Thus, one can trade-off the ease of programming with coherent memory for improved performance with non-coherent memory. As an example, I present a program to solve a linear system of equations using an asynchronous iterative algorithm. This program uses all the behaviors offered by Mermera. Third, I describe the implementation of Mermera on a BBN Butterfly TC2000 and on a network of workstations. The performance of a version of the equation solving program that uses all the behaviors of Mermera is compared with that of a version that uses coherent behavior only. For a system of 1000 equations the former exhibits at least a 5-fold improvement in convergence time over the latter. The version using coherent behavior only does not benefit from employing more than one workstation to solve the problem while the program using non-coherent behavior continues to achieve improved performance as the number of workstations is increased from 1 to 6. This measurement corroborates our belief that non-coherent shared memory can be a performance boon for some applications.

CSR: Constrained Selfish Routing in Ad-hoc Networks

Routing protocols for ad-hoc networks assume that the nodes forming the network are either under a single authority, or else that they would be altruistically forwarding data for other nodes with no expectation of a return. These assumptions are unrealistic since in ad-hoc networks, nodes are likely to be autonomous and rational (selfish), and thus unwilling to help unless they have an incentive to do so. Providing such incentives is an important aspect that should be considered when designing ad-hoc routing protocols. In this paper, we propose a dynamic, decentralized routing protocol for ad-hoc networks that provides incentives in the form of payments to intermediate nodes used to forward data for others. In our Constrained Selfish Routing (CSR) protocol, game-theoretic approaches are used to calculate payments (incentives) that ensure both the truthfulness of participating nodes and the fairness of the CSR protocol. We show through simulations that CSR is an energy efficient protocol and that it provides lower communication overhead in the best and average cases compared to existing approaches.

Amorphous Placement and Retrieval of Sensory Data in Sparse Mobile Ad-Hoc Networks

Abstract—Personal communication devices are increasingly being equipped with sensors that are able to passively collect information from their surroundings – information that could be stored in fairly small local caches. We envision a system in which users of such devices use their collective sensing, storage, and communication resources to query the state of (possibly remote) neighborhoods. The goal of such a system is to achieve the highest query success ratio using the least communication overhead (power). We show that the use of Data Centric Storage (DCS), or directed placement, is a viable approach for achieving this goal, but only when the underlying network is well connected. Alternatively, we propose, amorphous placement, in which sensory samples are cached locally and informed exchanges of cached samples is used to diffuse the sensory data throughout the whole network. In handling queries, the local cache is searched first for potential answers. If unsuccessful, the query is forwarded to one or more direct neighbors for answers. This technique leverages node mobility and caching capabilities to avoid the multi-hop communication overhead of directed placement. Using a simplified mobility model, we provide analytical lower and upper bounds on the ability of amorphous placement to achieve uniform field coverage in one and two dimensions. We show that combining informed shuffling of cached samples upon an encounter between two nodes, with the querying of direct neighbors could lead to significant performance improvements. For instance, under realistic mobility models, our simulation experiments show that amorphous placement achieves 10% to 40% better query answering ratio at a 25% to 35% savings in consumed power over directed placement.

JTP: An Energy-conscious Transport Protocol for Wireless Ad Hoc Networks

Within a recently developed low-power ad hoc network system, we present a transport protocol (JTP) whose goal is to reduce power consumption without trading off delivery requirements of applications. JTP has the following features: it is lightweight whereby end-nodes control in-network actions by encoding delivery requirements in packet headers; JTP enables applications to specify a range of reliability requirements, thus allocating the right energy budget to packets; JTP minimizes feedback control traffic from the destination by varying its frequency based on delivery requirements and stability of the network; JTP minimizes energy consumption by implementing in-network caching and increasing the chances that data retransmission requests from destinations "hit" these caches, thus avoiding costly source retransmissions; and JTP fairly allocates bandwidth among flows by backing off the sending rate of a source to account for in-network retransmissions on its behalf. Analysis and extensive simulations demonstrate the energy gains of JTP over one-size-fits-all transport protocols.

An Implementation of Mermera: A Shared Memory System that Mixes Coherence with Non-coherence

Coherent shared memory is a convenient, but inefficient, method of inter-process communication for parallel programs. By contrast, message passing can be less convenient, but more efficient. To get the benefits of both models, several non-coherent memory behaviors have recently been proposed in the literature. We present an implementation of Mermera, a shared memory system that supports both coherent and non-coherent behaviors in a manner that enables programmers to mix multiple behaviors in the same program[HS93]. A programmer can debug a Mermera program using coherent memory, and then improve its performance by selectively reducing the level of coherence in the parts that are critical to performance. Mermera permits a trade-off of coherence for performance. We analyze this trade-off through measurements of our implementation, and by an example that illustrates the style of programming needed to exploit non-coherence. We find that, even on a small network of workstations, the performance advantage of non-coherence is compelling. Raw non-coherent memory operations perform 20-40~times better than non-coherent memory operations. An example application program is shown to run 5-11~times faster when permitted to exploit non-coherence. We conclude by commenting on our use of the Isis Toolkit of multicast protocols in implementing Mermera.

Real-Time Spatio-Temporal Query Processing in Mobile Ad-Hoc Sensor Networks

Personal communication devices are increasingly equipped with sensors that are able to collect and locally store information from their environs. The mobility of users carrying such devices, and hence the mobility of sensor readings in space and time, opens new horizons for interesting applications. In particular, we envision a system in which the collective sensing, storage and communication resources, and mobility of these devices could be leveraged to query the state of (possibly remote) neighborhoods. Such queries would have spatio-temporal constraints which must be met for the query answers to be useful. Using a simplified mobility model, we analytically quantify the benefits from cooperation (in terms of the system's ability to satisfy spatio-temporal constraints), which we show to go beyond simple space-time tradeoffs. In managing the limited storage resources of such cooperative systems, the goal should be to minimize the number of unsatisfiable spatio-temporal constraints. We show that Data Centric Storage (DCS), or "directed placement", is a viable approach for achieving this goal, but only when the underlying network is well connected. Alternatively, we propose, "amorphous placement", in which sensory samples are cached locally, and shuffling of cached samples is used to diffuse the sensory data throughout the whole network. We evaluate conditions under which directed versus amorphous placement strategies would be more efficient. These results lead us to propose a hybrid placement strategy, in which the spatio-temporal constraints associated with a sensory data type determine the most appropriate placement strategy for that data type. We perform an extensive simulation study to evaluate the performance of directed, amorphous, and hybrid placement protocols when applied to queries that are subject to timing constraints. Our results show that, directed placement is better for queries with moderately tight deadlines, whereas amorphous placement is better for queries with looser deadlines, and that under most operational conditions, the hybrid technique gives the best compromise.

Communicable Memory and Lazy Barriers for Bulk Synchronous Parallelism in BSPk

Communication and synchronization stand as the dual bottlenecks in the performance of parallel systems, and especially those that attempt to alleviate the programming burden by incurring overhead in these two domains. We formulate the notions of communicable memory and lazy barriers to help achieve efficient communication and synchronization. These concepts are developed in the context of BSPk, a toolkit library for programming networks of workstations|and other distributed memory architectures in general|based on the Bulk Synchronous Parallel (BSP) model. BSPk emphasizes efficiency in communication by minimizing local memory-to-memory copying, and in barrier synchronization by not forcing a process to wait unless it needs remote data. Both the message passing (MP) and distributed shared memory (DSM) programming styles are supported in BSPk. MP helps processes efficiently exchange short-lived unnamed data values, when the identity of either the sender or receiver is known to the other party. By contrast, DSM supports communication between processes that may be mutually anonymous, so long as they can agree on variable names in which to store shared temporary or long-lived data.

Measuring Bottleneck Bandwidth of Targeted Path Segments

Accurate measurement of network bandwidth is crucial for flexible Internet applications and protocols which actively manage and dynamically adapt to changing utilization of network resources. These applications must do so to perform tasks such as distributing and delivering high-bandwidth media, scheduling service requests and performing admission control. Extensive work has focused on two approaches to measuring bandwidth: measuring it hop-by-hop, and measuring it end-to-end along a path. Unfortunately, best-practice techniques for the former are inefficient and techniques for the latter are only able to observe bottlenecks visible at end-to-end scope. In this paper, we develop and simulate end-to-end probing methods which can measure bottleneck bandwidth along arbitrary, targeted subpaths of a path in the network, including subpaths shared by a set of flows. As another important contribution, we describe a number of practical applications which we foresee as standing to benefit from solutions to this problem, especially in emerging, flexible network architectures such as overlay networks, ad-hoc networks, peer-to-peer architectures and massively accessed content servers.

An Efficient User-Level Shared Memory Mechanism for Application-Specific Extensions

This paper focuses on an efficient user-level method for the deployment of application-specific extensions, using commodity operating systems and hardware. A sandboxing technique is described that supports multiple extensions within a shared virtual address space. Applications can register sandboxed code with the system, so that it may be executed in the context of any process. Such code may be used to implement generic routines and handlers for a class of applications, or system service extensions that complement the functionality of the core kernel. Using our approach, application-specific extensions can be written like conventional user-level code, utilizing libraries and system calls, with the advantage that they may be executed without the traditional costs of scheduling and context-switching between process-level protection domains. No special hardware support such as segmentation or tagged translation look-aside buffers (TLBs) is required. Instead, our ``user-level sandboxing'' mechanism requires only paged-based virtual memory support, given that sandboxed extensions are either written by a trusted source or are guaranteed to be memory-safe (e.g., using type-safe languages). Using a fast method of upcalls, we show how our mechanism provides significant performance improvements over traditional methods of invoking user-level services. As an application of our approach, we have implemented a user-level network subsystem that avoids data copying via the kernel and, in many cases, yields far greater network throughput than kernel-level approaches.

Distributed Parallel Computing in Mermera: Mixing Noncoherent Shared Memories

Programmers of parallel processes that communicate through shared globally distributed data structures (DDS) face a difficult choice. Either they must explicitly program DDS management, by partitioning or replicating it over multiple distributed memory modules, or be content with a high latency coherent (sequentially consistent) memory abstraction that hides the DDS' distribution. We present Mermera, a new formalism and system that enable a smooth spectrum of noncoherent shared memory behaviors to coexist between the above two extremes. Our approach allows us to define known noncoherent memories in a new simple way, to identify new memory behaviors, and to characterize generic mixed-behavior computations. The latter are useful for programming using multiple behaviors that complement each others' advantages. On the practical side, we show that the large class of programs that use asynchronous iterative methods (AIM) can run correctly on slow memory, one of the weakest, and hence most efficient and fault-tolerant, noncoherence conditions. An example AIM program to solve linear equations, is developed to illustrate: (1) the need for concurrently mixing memory behaviors, and, (2) the performance gains attainable via noncoherence. Other program classes tolerate weak memory consistency by synchronizing in such a way as to yield executions indistinguishable from coherent ones. AIM computations on noncoherent memory yield noncoherent, yet correct, computations. We report performance data that exemplifies the potential benefits of noncoherence, in terms of raw memory performance, as well as application speed.

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.

Using Speculation to Reduce Server Load and Service Time on the WWW

Speculative service implies that a client's request for a document is serviced by sending, in addition to the document requested, a number of other documents (or pointers thereto) that the server speculates will be requested by the client in the near future. This speculation is based on statistical information that the server maintains for each document it serves. The notion of speculative service is analogous to prefetching, which is used to improve cache performance in distributed/parallel shared memory systems, with the exception that servers (not clients) control when and what to prefetch. Using trace simulations based on the logs of our departmental HTTP server http://cs-www.bu.edu, we show that both server load and service time could be reduced considerably, if speculative service is used. This is above and beyond what is currently achievable using client-side caching [3] and server-side dissemination [2]. We identify a number of parameters that could be used to fine-tune the level of speculation performed by the server.

Fuzzy ART: Fast Stable Learning and Categorization of Analog Patterns by an Adaptive Resonance System

A Fuzzy ART model capable of rapid stable learning of recognition categories in response to arbitrary sequences of analog or binary input patterns is described. Fuzzy ART incorporates computations from fuzzy set theory into the ART 1 neural network, which learns to categorize only binary input patterns. The generalization to learning both analog and binary input patterns is achieved by replacing appearances of the intersection operator (n) in AHT 1 by the MIN operator (Λ) of fuzzy set theory. The MIN operator reduces to the intersection operator in the binary case. Category proliferation is prevented by normalizing input vectors at a preprocessing stage. A normalization procedure called complement coding leads to a symmetric theory in which the MIN operator (Λ) and the MAX operator (v) of fuzzy set theory play complementary roles. Complement coding uses on-cells and off-cells to represent the input pattern, and preserves individual feature amplitudes while normalizing the total on-cell/off-cell vector. Learning is stable because all adaptive weights can only decrease in time. Decreasing weights correspond to increasing sizes of category "boxes". Smaller vigilance values lead to larger category boxes. Learning stops when the input space is covered by boxes. With fast learning and a finite input set of arbitrary size and composition, learning stabilizes after just one presentation of each input pattern. A fast-commit slow-recode option combines fast learning with a forgetting rule that buffers system memory against noise. Using this option, rare events can be rapidly learned, yet previously learned memories are not rapidly erased in response to statistically unreliable input fluctuations.