Source-level global optimizations for fine-grain distributed shared memory systems

Distributed shared memory (DSM) systems allow programmers to use a shared memory programming model on distributed systems. However, DSM systems present new challenges to performance evaluation and debugging. While programmers write programs with a shared memory model using shared variables and shared memory synchronization primitives to coordinate processes, most performance problems arise from the implementation of these abstractions with messages. Because the DSM abstraction hides the underlying message passing, it is difficult to identify the sources of performance problems. We have developed an instrumentation system and performance evaluation techniques for DSM systems that correlate message passing behavior in the lower levels of a DSM system with the higher-level shared memory programming abstractions. We have developed a prototype tool called Xunify to demonstrate the feasibility of such an approach. Xunify provides graphical animations of DSM programs at three distinct levels and offers a powerful event filtering facility capable to help show system behavior corresponding to specific DSM constructs.

Distributed Shared Memory (DSM) systems have gained popular acceptance by combining the scalability and low cost of distributed system with the ease of use of single address space. Many new hardware DSM and software DSM systems have been proposed in recent years. In general, benchmarking is widely used to demonstrate the performance advantages of new systems. However, the common method used to summarize the measured results is the arithmetic mean of ratios, which is incorrect in some cases. Furthermore, many published papers list a lot of data only, and do not summarize them effectively, which confuse users greatly. In fact, many users want to get a single number as conclusion, which is not provided in old summarizing techniques. Therefore, a new data-summarizing technique based on confidence interval is proposed in this paper. The new technique includes two data-summarizing methods: (1) paired confidence interval method; (2) unpaired confidence interval method. With this new technique, it is concluded that at some confidence one system is better than others. Four examples are shown to demonstrate the advantages of this new technique. Furthermore, with the help of confidence level, it is proposed to standardize the benchmarks used for evaluating DSM systems so that a convincing result can be got. In addition, the new summarizing technique fits not only for evaluating DSM systems, but also for evaluating other systems, such as memory system and communication systems.

Shared virtual memory (SVM) systems, because of their software implementation, enable shared-memory programming at a low design and maintenance cost. Nevertheless, as hardware implementations become faster, their performance is still far from that achieved by distributed shared memory (DSM) systems. Nowadays, SVM systems use relaxed memory consistency models and multiple writer protocols as techniques to reduce latencies and false sharing, respectively. However, these techniques induce additional overhead that decreases system performance. We performed a study of workload behavior aimed at improving the design of SVM protocols. The work focused on the identification of the type of shared data patterns that can appear in the accesses to protected sections using semaphores. Most coherence actions in SVM systems are performed as a consequence of the write operations executed in critical sections, so we pay special attention to the write operations performed when multiple writers are allowed. As these write operations may present spatial locality, we also study the write patterns on shared pages with similar behaviour. Different software filters are applied in the instrumented parallel workloads selected to capture and classify the most common sharing patterns. This enables the recognition of those patterns in which coherence overhead can be reduced by modifying the coherence actions performed by the protocol. Despite the fact that the performance evaluation of new coherence solutions is not our main goal, the ideas presented to improve the behaviour of SVM systems can be implemented at a reasonable hardware/software cost.

Nowadays clusters are one of the most used platforms in High Performance Computing and most programmers use the Message Passing Interface (MPI) library to program their applications in these distributed platforms getting their maximum performance, although it is a complex task. On the other side, OpenMP has been established as the de facto standard to program applications on shared memory platforms because it is easy to use and obtains good performance without too much effort. So, could it be possible to join both worlds? Could programmers use the easiness of OpenMP in distributed platforms? A lot of researchers think so. And one of the developed ideas is the distributed shared memory (DSM), a software layer on top of a distributed platform giving an abstract shared memory view to the applications. Even though it seems a good solution it also has some inconveniences. The memory coherence between the nodes in the platform is difficult to maintain (complex management, scalability issues, high overhead and others) and the latency of the remote-memory accesses which can be orders of magnitude greater than on a shared bus due to the interconnection network. Therefore this research improves the performance of OpenMP applications being executed on distributed memory platforms using a DSM with sequential consistency evaluating thoroughly the results from the NAS parallel benchmarks. The vast majority of designed DSMs use a relaxed consistency model because it avoids some major problems in the area. In contrast, we use a sequential consistency model because we think that showing these potential problems that otherwise are hidden may allow the finding of some solutions and, therefore, apply them to both models. The main idea behind this work is that both runtimes, the OpenMP and the DSM layer, should cooperate to achieve good performance, otherwise they interfere one each other trashing the final performance of applications. We develop three different contributions to improve the performance of these applications: (a) a technique to avoid false sharing at runtime, (b) a technique to mimic the MPI behaviour, where produced data is forwarded to their consumers and, finally, (c) a mechanism to avoid the network congestion due to the DSM coherence messages. The NAS Parallel Benchmarks are used to test the contributions. The results of this work shows that the false-sharing problem is a relative problem depending on each application. Another result is the importance to move the data flow outside of the critical path and to use techniques that forwards data as early as possible, similar to MPI, benefits the final application performance. Additionally, this data movement is usually concentrated at single points and affects the application performance due to the limited bandwidth of the network. Therefore it is necessary to provide mechanisms that allows the distribution of this data through the computation time using an otherwise idle network. Finally, results shows that the proposed contributions improve the performance of OpenMP applications on this kind of environments.

The management of memory coherence is an important problem in distributed shared memory (DSM) system. In a cache-based coherence DSM system using linked list structure, the key to maintaining the coherence and improving system performance is how to manage the owner in the linked list. This paper presents the design of a new management protocol—NONH (New-Owner New-Head) and its performance evaluation. The analysis results show that this protocol can improve the scalability and performence of a coherent DSM system using linked list It is also suitable for managing the cache coherency in tree-like hierarchical architecture.

This paper presents the design and implementation of our novel home-based software Distributed Shared Memory (DSM) system. We call our system Lazy Home-based Scope Consistency (LHScC) since the system implements scope consistency lazily based on the home-based protocol. Our technique is different from others in that an optimum home node is decided before updating a home node. The performance evaluation shows that LHScC is more scalable than other DSM systems and removes the home assignment problem in the usual home-based protocol.

Software based Distributed Shared Memory (DSM) systems have been the focus of considerable research effort, primarily in improving performance and consistency protocols. Unfortunately, computer clusters present a number of challenges for any DSM systems that are not solvable through consistency protocols alone. These challenges relate to the ability of DSM systems to adjust to load fluctuations, computers being added/removed from the cluster, to deal with faults, and the ability to use DSM objects larger than the available physical memory. We present here a proposal for the Synergy Distributed Shared Memory System and its integration with the virtual memory, group communication and process migration services of the Genesis Cluster Operating System.

The performance of page-based software shared virtual memory (SVM) is still far from that achieved on hardware-coherent distributed shared memory (DSM) systems. The interrupt cost for asynchronous protocol processing has been found to be a key source of performance loss and complexity.This paper shows that by providing simple and general support for asynchronous message handling in a commodity network interface (NI), and by altering SVM protocols appropriately, protocol activity can be decoupled from asynchronous message handling and the need for interrupts or polling can be eliminated. The NI mechanisms needed are generic, not SVM-dependent. They also require neither visibility into the node memory system nor code instrumentation to identify memory operations. We prototype the mechanisms and such a synchronous home-based LRC protocol, called GeNIMA (GEneral-purpose Network Interface support in a shared Memory Abstraction), on a cluster of SMPs with a programmable NI, though the mechanisms are simple and do not require programmability.We find that the performance improvements are substantial, bringing performance on a small-scale SMP cluster much closer to that of hardware-coherent shared memory for many applications, and we show the value of each of the mechanisms in different applications. Application performance improves by about 37% on average for reasonably well performing applications, even on our relatively slow programmable NI, and more for others. We discuss the key remaining bottlenecks at the protocol level and use a firmware performance monitor in the NI to understand the interactions with and the implications for the communication layer.

A coherent global address space in a distributed system enables shared memory programming in a much larger scale than a single multicore or a single SMP. Without dedicated hardware support at this scale, the solution is a software distributed shared memory (DSM) system. However, traditional approaches to coherence (centralized via active home-node directories) and critical-section execution (distributed across nodes and cores) are inherently unfit for such a scenario. Instead, it is crucial to make decisions locally and avoid the long latencies imposed by both network and software message handlers. Likewise, synchronization is fast if it rarely involves communication with distant nodes (or even other sockets). To minimize the amount of long-latency communication required in both coherence and critical section execution, we propose a DSM system with a novel coherence protocol, and a novel hierarchical queue delegation locking approach. More specifically, we propose an approach, suitable for Data-Race-Free programs, based on self-invalidation, self-downgrade, and passive data classification directories that require no message handlers, thereby incurring no extra latency. For fast synchronization we extend Queue Delegation Locking to execute critical sections in large batches on a single core before passing execution along to other cores, sockets, or nodes, in that hierarchical order. The result is a software DSM system called Argo which localizes as many decisions as possible and allows high parallel performance with little overhead on synchronization when compared to prior DSM implementations.

Directory-based write-invalidate cache coherence protocols have been widely used in distributed shared memory (DSM) systems, in which cache invalidation overheads occupy a large part of the system overheads. In this paper, we propose a novel tree-based multidestination multicast scheme TBM which involves a new efficient multidestination message format. TBM combines the best features of two existing approaches: tree-based multicast and multidestination message passing. With the new scheme, only one invalidation message and less than 2/sup [log2(n+1)]-1/ acknowledgement messages are required in one cache invalidation transaction when n processor nodes have the copies of the cache block. Detailed analysis and simulation in 2D mesh show that TBM is preferable to traditional Umesh, Hamiltonian Path and BRCP-HL multicast schemes, which indicates that current and future DSM systems can take advantage of this scheme to deliver better performance.

One of the factors that can influence the performance of a DSM system is the efficiency of multipoint access on the interconnection network. This work presents an evaluation of the Brazos system, a software implemented distributed shared memory (DSM) system designed for x86 SMP nodes running Windows NT and that takes advantage of multicast communication, with regard to different configurations of the communication network: broadcast fast Ethernet, switched fast Ethernet and LANE-ATM. In this evaluation, although the bandwidth of the ATM is larger than the fast Ethernet, low performance using LANE-ATM was observed, compared to the Brazos system performance using broadcast or switched Fast Ethernet. It shows the importance of the communication subsystem and the necessity of improvements on aspects like multipoint access, which has been not much explored on networks like ATM.

On the interconnection of causal memory systems

Data prefetching is a technique where a processing unit issues one or more non-blocking load operations before the actual data items are required. The access latency of prefetching can be alleviated by overlapping it with other executions which are independent of the prefetched data. In distributed shared memory (DSM) systems, remote memory accesses take much longer than local ones, and hence data prefetching should be effective for such systems. However, to our knowledge, relatively little research has been done for data prefetching on DSM systems. This paper is concerned with issues of supporting data prefetching on DSM systems. Our approach is to develop a new memory consistency semantic (MCS) model under which the prefetchable shared data objects, as well as the best moment to launch a prefetching operation, can be easily identified. Our new MCS, called aggressive consistency, utilizes the coherence-on-demand concept and supports a special synchronization operation called SYNC, which also acts as the prefetching indicator. Preliminary simulation results show that our prefetching approach, combined with aggressive consistency, can substantially improve the performance of DSM systems.

Despite the fast evolution of Distributed Shared Memory (DSM) systems, the load balance problem has not received enough attentions. This problem arises naturally after multithreading was introduced to DSM systems few years ago. The cognizance of it would bring us a significant improvement in system performance. In this paper, we address it by proposing and experimentally evaluating a load balancing method called Dependence-Driven Load Balancing (DDLB) that is dedicated for multithreaded DSM systems. The most attractive feature of this method is to take thread dependence into account in making decisions for migration. In contrast to existing thread scheduling works which for the most part have relied on the information supplied before execution, we require no a priori information. Typically, DDLB embraces three policies, i.e. transfer policy, location policy and selection policy, and applies affinity scheduling. Finally, from the experimental results, the performance of the system with load balancing is improving evidently.

Two most commonly used classifications of reference locality are: temporal locality and spatial locality. This paper introduces a new class of reference locality, called Regional Locality, which is the program behavior that a set of addresses which are accessed in one critical or non-critical region will be very likely accessed as a whole in the same critical region or other non-critical regions. We proposed three updates propagation protocols based on Regional Locality in Distributed Shared Memory systems. These protocols include: Selective Lazy/Eager Updates Propagation protocol, First Hit Updates Propagation protocol, and Second Hit Updates Propagation protocol. Our experimental results indicate that Regional Locality exists in executions of many Distributed Shared Memory concurrent programs. We have shown that the proposed protocols outperform the existing updates propagation protocols based on temporal locality. Exploring Regional Locality in other shared memory systems would be an interesting future research direction