OpenMP‐oriented applications for distributed shared memory architectures

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 distributed shared memory (DSM) architecture is considered to be one of the most likely parallel computing environment candidate for the near future because of its ease of system scalability and facilitation for parallel programming. However, a naive program based on shared memory execution on a DSM machine often deteriorates performance, because of the overhead involved for maintaining cache coherency particularly with frequent remote memory accesses. We show that careful buffer management of parallel join processing on DSM can produce considerable performance improvements in comparison with a naive implementation. We propose four buffer management strategies for parallel hash join processing on the DSM architecture and actually implement them on the HP Exemplar SPP 1600. The basic strategy is to begin with the hash join algorithm for the shared everything architecture and then to consider the memory locality of DSM by distributing the hash table and data pool buffers among the nodes. The results of four buffering strategies are analyzed in detail. Consequently, we can conclude that, in order to achieve high performance on a DSM machine, our buffer management strategy in which the memory access pattern is extracted and buffers are allocated in the local memory of nodes to minimize memory access cost is very efficient.

High speed networks and rapidly improving microprocessor performance make the network of workstations an extremely important tool for parallel computing in order to speedup the execution of scientific applications. Shared memory is an attractive programming model for designing parallel and distributed applications, where the programmer can focus on algorithmic development rather than data partition and communication. Based on this important characteristic, the design of systems to provide the shared memory abstraction on physically distributed memory machines has been developed, known as Distributed Shared Memory (DSM). DSM is built using specific software to combine a number of computer hardware resources into one computing environment. Such an environment not only provides an easy way to execute parallel applications, but also combines available computational resources with the purpose of speeding up execution of these applications. DSM systems need to maintain data consistency in memory, which usually leads to communication overhead. Therefore, there exists a number of strategies that can be used to overcome this overhead issue and improve overall performance. Strategies as prefetching have been proven to show great performance in DSM systems, since they can reduce data access communication latencies from remote nodes. On the other hand, these strategies also transfer unnecessary prefetching pages to remote nodes. In this research paper, we focus on the access pattern during execution of a parallel application, and then analyze the data type and behavior of parallel applications. We propose an adaptive data classification scheme to improve prefetching strategy with the goal to improve overall performance. Adaptive data classification scheme classifies data according to the accessing sequence of pages, so that the home node uses past history access patterns of remote nodes to decide whether it needs to transfer related pages to remote nodes. From experimental results, we can observe that our proposed method can increase the accuracy of data access in effective prefetch strategy by reducing the number of page faults and misprefetching. Experimental results using our proposed classification scheme show a performance improvement of about 9---25% over the same benchmark applications running on top of an original JIAJIA DSM system.

Scalable 3D Hybrid Parallel Delaunay Image-to-Mesh Conversion Algorithm for Distributed Shared Memory Architectures

The paper introduces a software-only object based Distributed Shared Memory (DSM) implementation designed as an extension to the Microsoft .NET Framework, which adds facilities for sharing objects following the Multiple Readers Multiple Writer (MRMW) memory model, including object replication. The presented implementation is facilitated by a memory coherence protocol previously developed by the author, which uses group communication using IP multicasting and delivers causally consistent memory. The paper describes an attempt to implement the DSM wholly in the context of .NET Remoting by extending it with an UDP-based channel service, a proxy implementation for object access interception and threads performing the receiving and sending parts of the coherence protocol, as well as the final implementation approach. The described DSM implementation allows the construction of distributed applications with a simple programming model, that can be deployed in any networked environment where multicast is available.

Practical solutions are important in meeting the educational needs of patients and their caregivers. Using mobile technology offers a tremendous opportunity to teach specific skills to patients and/or to caregivers. This study aimed to develop a mobile-based application to support patients and their caregivers in meeting the tracheal suctioning information needs of patients with tracheostomies. The "ASSURE Instructional Design Model" was used to develop the material. This model enables the mobile-based application to be developed stepwise, which consists of 6 stages. The designing and development process of the mobile-based application requires following each step carefully, collaborating with different teams, obtaining the expert opinions and learners' feedback, as well as revising the material until it reaches the desired form. This article reports the design and development process of the mobile-based application by using the ASSURE Model and provides guidelines for researchers willing to study in this field.

To solve problems that require far more memory than a single machine can supply, data can be swapped to disk in some manner, it can be compressed, and/or the memory of multiple parallel machines can be used to provide enough memory and storage space. Instead of implementing either functionality anew and specific for each application, or instead of relying on the operating system’s swapping algorithms (which are inflexible, not algorithm-aware, and often limited in their fixed storage capacity), our solution is a Large Virtual Machine (LVM) that transparently provides a large address space to applications and that is more flexible and efficient than operating system approaches.LVM is a virtual machine for Java that is designed to support large address spaces for billions of objects. It swaps objects out to disk, compresses objects where needed, and uses multiple parallel machines in a Distributed Shared Memory (DSM) setting. The latter is the main focus of this paper. Allocation and collection performance is similar to well-known JVMs if no swapping is needed. With swapping and clustering, we are able to create a list containing 1.2×108 elements far faster than other JVMs. LVM’s swapping is up to 10 times faster than OS-level swapping. A swap-aware GC algorithm helps by a factor of 3.KeywordsVirtual MachineHash TableGarbage CollectionAddress SpaceGarbage CollectorThese keywords were added by machine and not by the authors. This process is experimental and the keywords may be updated as the learning algorithm improves.

Tachyon primary system which introduces recently is a high performance supercomputer that composed with AMD Barcelona nodes. In this paper, we will verify the performance and parallel scalability of TachyonIn by using multi-zone NAS Parallel Benchmark(NPB) which is one of a program with hybrid parallel method. To test performance of hybrid parallel execution, B and C classes of BT-MZ in NPB version 3.3 were used. And the parallel scalability test has finished with Tachyon`s 1024 processes. It is the first time in Korea to get a result of hybrid parallel computing calculation using more than 1024 processes. Hybrid parallel method in high performance computing system with multi-core technology like Tachyon describes that it can be very efficient and useful parallel performance benchmarks.

In this paper, we introduce an MPEG-V based haptic authoring tool intended for simplifying the development process of haptics-enabled multimedia applications. The developed tool provides a web-based interface for users to create haptic environments by importing 3D models and adding haptic properties to them. The user can then export the resulting environment to a standard MPEG-V format. The latter can be imported to a haptic player that renders the described haptics-enabled 3D scene. The proposed tool can support many haptic devices, including Geomagic Devices, Force Dimension Devices, Novint Falcon Devices, and Moog FCS HapticMaster Devices. We conduct a proof of concept HTML5 haptic game project and user studies on haptic effects, development process and user interface, which shows our tool's effectiveness in simplifying the development process of haptics-enabled multimedia applications.

Today, it is commonly accepted that speedup and efficiency are not granted automatically when developing or porting software for High Performance Computing (HPC) platforms. The reasons are manifold and actively investigated in the community. Software monitors are essential for these studies as they provide the raw performance data to be analyzed. We propose an online monitoring workflow for event-based performance analysis that takes into account the significant changes in system architecture towards the development of exascale supercomputers. Critical properties are: communication across large numbers of processing elements, limited I/O capabilities, and a decreasing memory-per-core ratio. We present a hierarchical data management and steering workflow that directly couples the application, measurement, and analysis processes, thus eliminating the need for extensive communication and buffering. The workflow is closely integrated with the native system communication API to enable best communication across processing elements. The memory issue is addressed with a new lossy hierarchical data compression technique for in-memory storage, intended for small, fixed-size buffers. Further, we abandon secondary storage to avoid potential I/O challenges. We demonstrate the feasibility of our design with a prototype implementation that features services for data collection, analysis, and runtime compression. Our evaluation extrapolates results obtained with the NAS Parallel Benchmarks at up to 2,048 processes to an exascale workflow.

The fundamental model of memory requires each read to an address to return the most recent value written to the same address. This model becomes a consistency requirement for parallel shared memory machines, and is taken as a necessary property for correct algorithm execution. In Distributed Shared Memory (DSM) systems the presence of processor memory caches means that the processors may develop historical and therefore different views of the distributed memory. The paper examines the relationship between synchronisation and consistency, proposes that synchronisation should enforce consistency, and examines the consequences using barriers as an example. A typical shared memory MIMD program is shown to execute correctly without normal consistency, and the experiment suggests that the relationship between synchronisation and consistency bears further investigation.

A hybrid parallel measurement system offers the potential to fuse the principal advantages of probe-based tools, with their exact measures of performance and ability to capture event semantics, and sampling-based tools, with their ability to observe performance detail with less overhead. Creating a hybrid profiling solution is challenging because it requires new mechanisms for integrating probe and sample measurements and calculating profile statistics during execution. In this paper, we describe a general hybrid parallel profiling tool that has been implemented in the TAU Performance System. Its generality comes from the fact that all of the features of the individual methods are retained and can be flexibly controlled when combined to address the measurement requirements for a particular parallel application. The design of the hybrid profiling approach is described and the implementation of the prototype in TAU presented. We demonstrate hybrid profiling functionality first on a simple sequential program and then show its use for several OpenMP parallel codes from the NAS Parallel Benchmark. These experiments also highlight the improvements in overhead efficiency made possible by hybrid profiling. A large-scale ocean modeling code based on OpenMP and MPI, MPAS-Ocean, is used to show how the TAU hybrid profiling tool can be effective at exposing performance-limiting behavior that would be difficult to identify otherwise.

Emerging heterogeneous architectures do not necessarily provide cache-coherent shared memory across all components of the system. Although there are good reasons for this architectural decision, it does provide programmers with a challenge. Several programming models and approaches are currently available, including explicit data movement for offloading computation to coprocessors, and treating coprocessors as distributed memory machines by using message passing. This paper examines the potential of distributed shared memory (DSM) for addressing this programming challenge. We discuss how our recently proposed DSM system and its memory consistency model maps to the heterogeneous node context, and present experimental results that highlight the advantages and challenges of this approach.

This paper presents preliminary efforts to develop compilation and execution environments that achieve performance portability of multilevel parallelization on hierarchical architectures. Using the NAS parallel benchmarks, we first illustrate the lack of portable performance on state-of-the-art scalable parallel systems despite the use of two portable programming models, MPI and OpenMP. Then we present a dynamic compilation and execution framework that provides the desired portability through the use of program slices. These slices are used to select the optimal program decomposition on each architecture. Currently, our framework uses a simple incremental algorithm, which effectively identifies single or multi-level program decompositions that maximize performance. This algorithm can be used as a rule of thumb for automatic multilevel parallelization. The effectiveness of the approach is demonstrated on the NAS benchmarks running on two architectural platforms.

The match rate overall and for applicant subgroups such as allopathic senior medical students and foreign-trained physicians in the National Resident Matching Program (NRMP) Main Residency Match has remained remarkably stable over time. This suggests that the Match's competitiveness has not undergone any major changes.1 Nevertheless, nearly every specialty has experienced a steady increase in the number of residency programs to which each student applies, creating waste in the residency recruitment process. As a result, more resources are being expended without demonstrable improvement in Match outcomes.2 The increased financial cost for the application process and in-person interviews burdens applicants, while residency programs devote more faculty and staff resources to review and interview a larger number of applicants. Given medical students' focus on recruitment and time away for interviews, the educational value of the fourth year—to ensure preparedness for residency—is eroding.3,4 Furthermore, the influx of applications may lead to program practices that rely on easily filterable markers of success, such as United States Medical Licensing Examination (USMLE) scores. Despite genuine interest and even prospects of being a “good fit” based on other metrics, many applicants may be disadvantaged in the current system, because highly competitive applicants receive more interview invitations than they can accept and may horde interview slots until later in the recruitment season. Students and residency program directors alike lament the phenomenon of application inflation, yet the numbers continue to grow.Over the past 5 years, several solutions to the phenomenon of application inflation and its concomitant waste of resources for students and programs have been proposed. Some have called for limits to the total number of applications per applicant, which presupposes that any competitive advantage or disadvantage would be felt evenly across all applicants.2,5 However, given applicants' consumer rights to apply to as many programs as financially feasible, this option has failed to gain traction. Further, students with special circumstances (eg, couples matching, or an irregularity on their records such as a failed class or failed USMLE attempt) may require an exception to a limit or risk going unmatched. Otolaryngology programs now require a program-specific paragraph for each student application, to encourage students to consider their “fit” for a particular program and to deter application inflation. A decline in the number of applications occurred the first year of implementation, along with an increase in unmatched programs, which brings up concerns about this approach.6While a cap on applications is unlikely to become a reality, medical student groups have repeatedly called for programs to provide specific data about characteristics of matched applicants, to provide guidance about an applicant's competitiveness.7 Otherwise, how is a student to know if they are sufficiently competitive? Programs have been reluctant to publish these data; they worry that providing the data may limit opportunities to attract more competitive applicants. It appears that the undergraduate medical education (UME) and graduate medical education (GME) communities are at an impasse.In this issue of the Journal of Graduate Medical Education, Whipple and colleagues propose a potential solution, one that will not meaningfully disadvantage any applicant while improving the chances of students and programs to find the right match.8 Their proposed solution is elegantly simple: applicants would have the option to designate some programs (presumably a limited number) as “preferred programs.” The authors use real numbers from the 2014 Otolaryngology Match to demonstrate that, if medical students selected a subset of their applications as “preferred,” the majority of students would receive more interview invitations per application than in the current system. The authors demonstrate that only the most competitive applicants, evaluated by level of competitiveness in both “easy-to-assess” measures (USMLE scores, class rank, grade point average, Alpha Omega Alpha status, number of publications, geographic preference) and “hard-to-assess” measures (letters of recommendation, personal recommendations, personal statement, qualitative performance reviews, awards, volunteer activities, research interests), were ostensibly disadvantaged in this model. The most competitive applicants receive fewer invitations to interview in the new model. The most competitive students may receive more invitations than they can accommodate, and thus, a system in which students have the option to designate preferred programs would not actually disadvantage this group.The authors have proposed an intriguing solution. In their scenario, by knowing applicants' genuine interest, programs could devote more time to fully evaluate these applications for interview invitations. Program resources would be primarily spent on the review of those applicants who were sincerely interested in the program, thereby minimizing the waste inherent in the review of an applicant with less interest.However, there are key limitations to the proposed solution. First, it is not clear how student preference designation would perform in less competitive specialties and residency programs. The simulation model was performed in one of the most competitive surgical subspecialties—most specialty matches do not fit the otolaryngology model. Second, it may be difficult to garner support from students and medical schools for this system. While the authors propose that students would not be required to reveal preferences, students may feel pressured to reveal preferences or risk losing an interview invitation to a desired program. Third, by virtue of limiting the number of programs a student can designate as preferred (which presumes that programs will fill most or all of their interview slots with competitive applicants from the “preferred program” pool), the authors propose a system with a virtual cap rather than an actual one. To date, an application cap has not been acceptable to students or schools; therefore, it is not clear that a virtual cap would be any more welcome. Finally, students may see this as yet another example where they are being asked—essentially required—to disclose information while programs continue to resist publishing statistics of matched applicants. This imbalance will do nothing to improve the trust and communication between students and the programs to which they are applying for further training.The model proposed by Whipple and colleagues remains promising. To our knowledge, this is the first published argument to demonstrate that when residency program leaders know which applicants have sincere interest in their program, more interview invitations will be extended per applicant for the vast majority of students. However, while novel and creative, the proposed approach remains a technical fix to the challenge of application inflation. Essentially, it would redistribute invitations, but the reasoning is circular. Even if a greater share of applicants receives more invitations, it is not necessarily true that a better fit will result, applicants will be less stressed, or applicants or programs will be more satisfied with the process. Nor will this approach enable students to substantially refocus their energies on the educational value of their fourth year of medical school.We urge the medical education community to consider more radical solutions. The American Medical Association has recently called for proposals for institutions to pilot disruptive innovations across the UME to GME continuum—innovations that will promote students' preparedness for training and well-being in residency. This may create favorable conditions to explore new models for residency application, recruitment, and selection.In the era of competency-based education and promotion, and educational handoffs along the UME to GME continuum, perhaps the current residency application process is no longer the right “fit.” What if programs engaged in mission-driven recruitment and sought to attract applicants who have achieved competencies specific to this mission? Programs could list track(s), informed by stakeholder input and alumni outcomes, which reflect their strengths, along with the desired competencies achieved by incoming interns, for each track. An internal medicine program, for example, could choose to list any number of specializations: rural health, cardiology fellowship preparedness, health informatics, quality improvement, primary care, community practice, etc. Students and their medical schools would submit educational portfolios that reflect program “fit” rather than the current model that primarily emphasizes overall medical school performance and suitability for a core specialty. We could realize the vision of an actual educational continuum with an effective handoff across the UME to GME transition. Or, what if students were guaranteed residency positions in their institution (or a consortium of institutions) as part of their acceptance to medical school? They could then focus their energy on developing the competencies to be the best physicians they could be, rather than on developing a track record that will make them competitive for the residency match. Imagine that: medical students' tremendous capacities focused on becoming physicians rather than on being competitive residency applicants.