Single-node Systems Research Articles

Machine Learning Tool for Analyzing Finite Buffer Queueing Systems

Queueing delays are one very important performance measure for most engineering network systems. Providing low-delay systems is a major goal of service providers, as it is a leading concern for users/customers. These network systems and their performance measures are typically analyzed using queueing-based models. Even though there are several available strong and precise mathematical models for analyzing queueing systems, their applications are limited to simple and small-scale systems due to their lack of scalability in real-life systems. Researchers have spent a good portion of their efforts toward perfecting the analysis of such systems. Precise and accurate results are available for single-node systems with standard operations. However, for analyzing multi-node systems with complex operations, one has to resort to approximations or simulations. Some of these approximations usually give an oversimplified view of such systems; these approximations remain quite limited. In this paper, we present a machine learning tool that can potentially be used to analyze most finite buffer queues to obtain reasonable approximations for the mean number of items in such systems. The machine learning tool we develop is based on supervised learning using the Michaelis–Menten non-linear model used in biochemistry and the results are simple to obtain. It is fast and very scalable; these characteristics represent the main features of this approach compared to existing systems. The coefficient of determination R2 for all the examples presented are all higher than 90%, with some as high as 99.6%.

Controllability of Multilayer Networked Sampled-Data Systems.

The controllability analysis of networked systems is challenging due to their high dimensionality and complex structure. The influence of sampling on network controllability is rarely studied, making it an important topic to explore. In this article, the state controllability of multilayer networked sampled-data systems is studied, considering the deep network structure, multidimensional node dynamics, various inner couplings, and sampling patterns. Necessary and/or sufficient controllability conditions are proposed and validated by numerical and practical examples, requiring less computation than the classic Kalman criterion. Single-rate and multirate sampling patterns are analyzed, showing that adjusting the sampling rate of local channels can affect the controllability of the overall system. It is shown that the pathological sampling of single-node systems can be eliminated by an appropriate design of interlayer structures and inner couplings. In the case of systems with drive-response mode, the overall system may not lose controllability even when the response layer is uncontrollable. The results demonstrate that mutually coupled factors collectively affect the controllability of the multilayer networked sampled-data system.

A Deterministic Portable Parallel Pseudo-Random Number Generator for Pattern-Based Programming of Heterogeneous Parallel Systems

SkePU is a pattern-based high-level programming model for transparent program execution on heterogeneous parallel computing systems. A key feature of SkePU is that, in general, the selection of the execution platform for a skeleton-based function call need not be determined statically. On single-node systems, SkePU can select among CPU, multithreaded CPU, single or multi-GPU execution. Many scientific applications use pseudo-random number generators (PRNGs) as part of the computation. In the interest of correctness and debugging, deterministic parallel execution is a desirable property, which however requires a deterministically parallelized pseudo-random number generator. We present the API and implementation of a deterministic, portable parallel PRNG extension to SkePU that is scalable by design and exhibits the same behavior regardless where and with how many resources it is executed. We evaluate it with four probabilistic applications and show that the PRNG enables scalability on both multi-core CPU and GPU resources, and hence supports the universal portability of SkePU code even in the presence of PRNG calls, while source code complexity is reduced.

Evaluating Geospatial RDF Stores Using the Benchmark Geographica 2

Geospatial extensions of SPARQL, like GeoSPARQL and stSPARQL, have been defined since 2007, and while several geospatial RDF stores have implemented a substantial part of these extensions, other stores limited their support mostly on point geometry features. A parallel process with the above was that RDF frameworks evolved in an interesting way by presenting a more mature set of geospatial features, such as GeoSPARQL support and including the latest indexing technologies. As a logical consequence, a shift in the use of RDF frameworks is to be expected, from base platforms that users extend to create more complete geospatial RDF stores, to attractive finished RDF solutions for many geospatial applications. Alongside with the ever-increasing size of linked geospatial data that semantic stores need to handle, all the above provided our group the motivation to improve our single-node systems benchmark Geographica, originally defined in 2013. Geographica 2 is more comprehensive, because it now includes new geospatial RDF stores and frameworks, big real-world datasets of many hundred million triples with up to 50 million features of complex geometries, new tests and queries that reveal the scalability of these systems. The augmented and revised real-world workload of Geographica 2 tests the efficiency of primitive spatial functions in RDF stores, their performance in the geocoding scenario against the new Census dataset in addition to many other real use case scenarios and finally includes computation of statistics for geospatial datasets. A more detailed and systematic evaluation is performed using the synthetic workload. The new scalability workload aims at discovering the limits of centralized geospatial RDF stores of various architectures. It employs a set of six well-balanced real-world datasets with highly complex geometries covering many European countries and compares three RDF stores in terms of storage space, bulk loading and query response time. In addition, a special version of the benchmark has been created for systems with limited geospatial functionality and two more systems of this category are introduced along the six systems of the main benchmark, all stressed against point-only subsets of the workloads. Three out of the eight systems use an RDBMS for the persistence layer, while some of them offer a variety of persistence options.

Multinodal Acoustic Trapping Enables High Capacity and High Throughput Enrichment of Extracellular Vesicles and Microparticles in miRNA and MS Proteomics Studies.

We report a new design of an acoustophoretic trapping device with significantly increased capacity and throughput, compared to current commercial acoustic trapping systems. Acoustic trapping enables nanoparticle and extracellular vesicle (EV) enrichment without ultracentrifugation. Current commercial acoustic trapping technology uses an acoustic single-node resonance and typically operates at flow rates <50 μL/min, which limits the processing of the larger samples. Here, we use a larger capillary that supports an acoustic multinode resonance, which increased the seed particle capacity 40 times and throughput 25–40 times compared to single-node systems. The resulting increase in capacity and throughput was demonstrated by isolation of nanogram amounts of microRNA from acoustically trapped urinary EVs within 10 min. Additionally, the improved trapping performance enabled isolation of extracellular vesicles for downstream mass spectrometry analysis. This was demonstrated by the differential protein abundance profiling of urine samples (1–3 mL), derived from the non-trapped versus trapped urine samples.

Scalable and efficient graph traversal on high-throughput cluster

Graph is one of the most important data structures in modern big data applications and is widely used in various fields. Among many graph algorithms, the Breadth-First Search (BFS) algorithm is a classic algorithm to solve the graph traversal problem and also the key kernel of Graph500 benchmark. On modern CPU architecture, the implementation of graph traversal on single-node systems has achieved significant improvement. However, due to the low resource utilization and high communications overhead, graph traversal on distributed clusters suffers from poor performance and energy inefficiency. High-throughput cluster (HTCs) adopt High-Throughput many-core architecture, which has the characteristics of high concurrency, strong real-time, and low-power consumption. In this work, we propose several techniques, including asynchronous virtual ring method, thread caching scheme and vertex ID reordering to solve above problems and improve BFS performance on HTCs. We systematically evaluate optimized BFS algorithm and achieve 249.74 giga-traversed edges per second (GTEPS) on 72 nodes (2880 cores) HTCs. Compared with results on Graph500 list, the optimized algorithm achieves the highest node efficiency under the same cluster scale and the performance shows weakly linear scalability as the number of cluster nodes increases. With regard to efficiency, the average performance on HTCs is 3.47 GTEPS/node, which is the best among CPU-based distributed systems on the November 2019 Graph500 list.

Advancing model calibration and uncertainty analysis of SWAT models using cloud computing infrastructure: LCC-SWAT

Abstract Calibration and uncertainty analysis of a complex, over-parameterized environmental model such as the Soil and Water Assessment Tool (SWAT) requires thousands of simulation runs and multiple calibration iterations. A parallel calibration system is thus desired that can be deployed on cloud-based architectures for reducing calibration runtime. This paper presents a cloud-based calibration and uncertainty analysis system called LCC-SWAT that is designed for SWAT models. Two optimization techniques, sequential uncertainty fitting (SUFI-2) and dynamically dimensioned search (DDS), have been implemented in LCC-SWAT. Moreover, the cloud-based system has been deployed on the Southern Ontario Smart Computing Innovation Platform's (SOSCIP) Cloud Analytics platform for diagnostic assessment of parallel calibration runtime on both single-node and multi-node CPU architectures. Unlike other calibrations/uncertainty analysis systems developed on the cloud, this system is capable of generating a comprehensive set of statistical information automatically, which facilitates broader analyses of the performance of the SWAT models. Experimental results on SWAT models of different complexities showed that LCC-SWAT can reduce runtime significantly. The runtime reduction is more pronounced for more complex and computationally intensive models. However, the reported runtime efficiency is significantly higher for single node systems. Comparative experiments with DDS and SUFI-2 show that parallel DDS outperforms parallel SUFI-2 in terms of both parameter identifiability and reducing uncertainty in model simulations. LCC-SWAT is a flexible calibration system and other optimization algorithms and asynchronous parallelization strategies can be added to it in future.

Diffusion limits for networks of Markov-modulated infinite-server queues

This paper studies the diffusion limit for a network of infinite-server queues operating under Markov modulation, meaning that the system’s parameters depend on an autonomously evolving Markov chain, called the background process. In previous papers on single-node queues with Markov modulation, two variants were distinguished. In the first variant the arrival rate and the server speed are modulated, whereas in the second variant the arrival rate and the service requirement are modulated. The setup of the present paper, however, is more general: we not only extend single-node systems to a network setting, but also allow both the server speed and the service requirement to depend on the background process. For this model we derive a Functional Central Limit Theorem. In particular, we show that, after accelerating the arrival processes and the background process, a centered and normalized version of the network population vector converges to a multidimensional Ornstein–Uhlenbeck process. The proof of this result relies on weak convergence of stochastic integrals as well as continuous-mapping arguments.

Самая быстрая и энергоэффективная реализация алгоритма поиска в ширину на одноузловых различных параллельных архитектурах согласно рейтингу Graph500

The breadth-first search algorithm is one of the basic algorithms for graph traversing and is used in many other algorithms of high-level graph analysis. BFS is characterized with intensive irregular memory accesses and a random data dependency, which greatly complicates its parallelization to all existing architectures. The paper considers the implementation of the BFS algorithm (the core benchmark of the Graph500 rating) for processing large graphs on different architectures: Intel x86, IBM Power8+, Intel KNL and NVidia GPU. Algorithms for implementing breadth-first search will be considered, such as top-down traverse, bottom-up traverse, and a hybrid traverse that includes both top-down and bottom-up traverses. The features of the algorithm implementation on shared memory will be shown, as well as graph transformations (local sorting of graph vertices, global sorting of graph vertices, renumbering of all graph vertexes, compressed representation of graph vertices), which allow achieving the best performance and energy-efficiency in this algorithm among all single-node systems of Graph500 and GreenGraph500 ratings.

Software-based single-node multi-GPU systems for interactive display wall

A display wall, one of the multiple display environments (MDEs), has been widely adopted in various fields including consumer electronics. Conventional display walls are constructed by utilizing cluster-based systems where multiple displays are connected via multiple cluster nodes. However, cluster-based systems have suffered from communication overheads among the cluster nodes when they provide interaction features and run diverse applications. This paper proposes a software-based single-node multi-GPU system for interactive display walls. The proposed system allows multiple GPUs in a single-node system to connect multiple displays, by providing an individual compositor for each GPU in the system. This paper also analyzes memory management models of multi-GPU systems and compositing overheads for multiple displays. Experimental results show that the FPS (frames per second) of the proposed system outperforms that of the conventional system by up to 1.36x, reducing copy overheads among the GPUs by efficient memory management and distributing compositing overheads to multiple GPUs.

Interference Management for Distributed Parallel Applications in Consolidated Clusters

Consolidating multiple applications on a system can improve the overall resource utilization of data center systems. However, such consolidation can adversely affect the performance of some applications due to interference caused by resource contention. Despite many prior studies on the interference effects in single-node systems, the interference behaviors of distributed parallel applications have not been investigated thoroughly. With distributed applications, a local interference in a node can affect the whole execution of an application spanning many nodes. This paper studies an interference modeling methodology for distributed applications to predict their performance under interference effects in consolidated clusters. This study first characterizes the effects of interference for various distributed applications over different interference settings, and analyzes how diverse interference intensities on multiple nodes affect the overall performance. Based on the characterization, this study proposes a static profiling-based model for interference propagation and heterogeneity behaviors. In addition, this paper presents use case studies of the modeling method, two interference-aware placement techniques for consolidated virtual clusters, which attempt to maximize the overall throughput or to guarantee the quality-of-service.

Interference Management for Distributed Parallel Applications in Consolidated Clusters

Consolidating multiple applications on a system can improve the overall resource utilization of data center systems. However, such consolidation can adversely affect the performance of some applications due to interference caused by resource contention. Despite many prior studies on the interference effects in single-node systems, the interference behaviors of distributed parallel applications have not been investigated thoroughly. With distributed applications, a local interference in a node can affect the whole execution of an application spanning many nodes. This paper studies an interference modeling methodology for distributed applications to predict their performance under interference effects in consolidated clusters. This study first characterizes the effects of interference for various distributed applications over different interference settings, and analyzes how diverse interference intensities on multiple nodes affect the overall performance. Based on the characterization, this study proposes a static profiling-based model for interference propagation and heterogeneity behaviors. In addition, this paper presents use case studies of the modeling method, two interference-aware placement techniques for consolidated virtual clusters, which attempt to maximize the overall throughput or to guarantee the quality-of-service.

HadoopCL2: Motivating the Design of a Distributed, Heterogeneous Programming System With Machine-Learning Applications

Machine learning (ML) algorithms have garnered increased interest as they demonstrate improved ability to extract meaningful trends from large, diverse, and noisy data sets. While research is advancing the state-of-the-art in ML algorithms, it is difficult to drastically improve the real-world performance of these algorithms. Porting new and existing algorithms from single-node systems to multi-node clusters, or from architecturally homogeneous systems to heterogeneous systems, is a promising optimization technique. However, performing optimized ports is challenging for domain experts who may lack experience in distributed and heterogeneous software development. This work explores how challenges in ML application development on heterogeneous, distributed systems shaped the development of the HadoopCL2 (HCL2) programming system. ML applications guide this work because they exhibit features that make application development difficult: large & diverse datasets, complex algorithms, and the need for domain-specific knowledge. The goal of this work is a general, MapReduce programming system that outperforms existing programming systems. This work evaluates the performance and portability of HCL2 against five ML applications from the Mahout ML framework on two hardware platforms. HCL2 demonstrates speedups of greater than 20x relative to Mahout for three computationally heavy algorithms and maintains minor performance improvements for two I/O bound algorithms.

Research and implementation of a distributed transaction processing middleware

Currently, increasingly transactional requests require high-performance transaction processing systems as support. The performance of a distributed transaction processing system is superior to that of traditional single-node transaction processing system, and the characteristic of multi-node determines that distributed transaction processing systems should pay more attention to availability. For example, in traditional single-node systems, the performance of Berkeley DB is high, but its shortcoming of not supporting parallel writing across multiple nodes is weakening its availability and scalability in the distributed environment. This paper has designed and implemented a middleware-level distributed transaction processing system called POST, including a distributed database system called POSTBOX which is based on Berkeley DB and data partition, and a distributed transaction processing middleware called POSTMAN. POSTBOX inherits the availability of highly available Berkeley DB, and expands it with data partition. By Partition Replication Body (PRB), POSTBOX overcomes the native weakness of highly available Berkeley DB, which indicates that highly available Berkeley DB does not support parallel writing across multiple nodes; POSTMAN is a middleware adapting PRB. POSTMAN monitors POSTBOX in real-time via Partition Replication Body State Array (PRBSA), and ensures the correctness of transaction processing and transactions migration in the case of node failure. The actual test results show that POST possesses high availability, and has an obvious improvement of write performance compared with highly available Berkeley DB.

Scalable logging through emerging non-volatile memory

Emerging byte-addressable, non-volatile memory (NVM) is fundamentally changing the design principle of transaction logging. It potentially invalidates the need for flush-before-commit as log records are persistent immediately upon write. Distributed logging---a once prohibitive technique for single node systems in the DRAM era---becomes a promising solution to easing the logging bottleneck because of the non-volatility and high performance of NVM. In this paper, we advocate NVM and distributed logging on multicore and multi-socket hardware. We identify the challenges brought by distributed logging and discuss solutions. To protect committed work in NVM-based systems, we propose passive group commit , a lightweight, practical approach that leverages existing hardware and group commit. We expect that durable processor cache is the ultimate solution to protecting committed work and building reliable, scalable NVM-based systems in general. We evaluate distributed logging with logging-intensive workloads and show that distributed logging can achieve as much as ~3x speedup over centralized logging in a modern DBMS and that passive group commit only induces minuscule overhead.

Design and implementation of GXP make — A workflow system based on make

This paper describes a rationale behind designing workflow systems based on the Unix make by showing a number of idioms useful for workflows comprising many tasks. It also demonstrates a specific design and implementation of such a workflow system called GXP make. GXP make supports all the features of GNU make and extends its platforms from single node systems to clusters, clouds, supercomputers, and distributed systems. Interestingly, it is achieved by a very small code base that does not modify GNU make implementation at all. While not being ideal for performance, it achieved a useful performance and scalability of dispatching one million tasks in approximately 5000 s (200 tasks per second, including dependence analysis) on an 8 core Intel Nehalem node. For real applications, recognition and classification of protein–protein interactions from biomedical texts on a supercomputer with more than 8000 cores are described.

Generalized boson algebra and its entangled bipartite coherent states

Starting with a given generalized boson algebra known as the bosonized version of the quantum super-Hopf algebra, we employ the Hopf duality arguments to provide the dually conjugate function algebra . With both the Hopf algebras being finitely generated, we produce a closed form expression of the universal matrix that caps the duality and generalizes the familiar exponential map relating a Lie algebra with its corresponding group. Subsequently, using an inverse Mellin transform approach, the coherent states of single-node systems subject to the symmetry are found to be complete with a positive-definite integration measure. The nonclassical coalgebraic structure of the algebra is found to generate naturally entangled coherent states in bipartite composite systems.

Devirtualizable virtual machines enabling general, single-node, online maintenance

Maintenance is the dominant source of downtime at high availability sites. Unfortunately, the dominant mechanism for reducing this downtime, cluster rolling upgrade, has two shortcomings that have prevented its broad acceptance. First, cluster-style maintenance over many nodes is typically performed a few nodes at a time, mak-ing maintenance slow and often impractical. Second, cluster-style maintenance does not work on single-node systems, despite the fact that their unavailability during maintenance can be painful for organizations. In this paper, we propose a novel technique for online maintenance that uses virtual machines to provide maintenance on single nodes, allowing parallel maintenance over multiple nodes, and online maintenance for standalone servers. We present the Microvisor, our prototype virtual machine system that is custom tailored to the needs of online maintenance. Unlike general purpose virtual machine environments that induce continual 10-20% over-head, the Microvisor virtualizes the hardware only during periods of active maintenance, letting the guest OS run at full speed most of the time. Unlike past attempts at virtual machine optimization, we do not compromise OS transparency. We instead give up generality and tailor our virtual machine system to the minimum needs of online maintenance, eschewing features, such as I/O and memory virtualization, that it does not strictly require. The result is a very thin virtual machine system that induces only 5.6% CPU overhead when virtualizing the hardware, and zero CPU overhead when devirtualized . Using the Microvisor, we demonstrate an online OS upgrade on a live, single-node web server, reducing downtime from one hour to less than one minute.

Devirtualizable virtual machines enabling general, single-node, online maintenance

Maintenance is the dominant source of downtime at high availability sites. Unfortunately, the dominant mechanism for reducing this downtime, cluster rolling upgrade, has two shortcomings that have prevented its broad acceptance. First, cluster-style maintenance over many nodes is typically performed a few nodes at a time, mak-ing maintenance slow and often impractical. Second, cluster-style maintenance does not work on single-node systems, despite the fact that their unavailability during maintenance can be painful for organizations. In this paper, we propose a novel technique for online maintenance that uses virtual machines to provide maintenance on single nodes, allowing parallel maintenance over multiple nodes, and online maintenance for standalone servers. We present the Microvisor, our prototype virtual machine system that is custom tailored to the needs of online maintenance. Unlike general purpose virtual machine environments that induce continual 10-20% over-head, the Microvisor virtualizes the hardware only during periods of active maintenance, letting the guest OS run at full speed most of the time. Unlike past attempts at virtual machine optimization, we do not compromise OS transparency. We instead give up generality and tailor our virtual machine system to the minimum needs of online maintenance, eschewing features, such as I/O and memory virtualization, that it does not strictly require. The result is a very thin virtual machine system that induces only 5.6% CPU overhead when virtualizing the hardware, and zero CPU overhead when devirtualized . Using the Microvisor, we demonstrate an online OS upgrade on a live, single-node web server, reducing downtime from one hour to less than one minute.

Devirtualizable virtual machines enabling general, single-node, online maintenance

Maintenance is the dominant source of downtime at high availability sites. Unfortunately, the dominant mechanism for reducing this downtime, cluster rolling upgrade, has two shortcomings that have prevented its broad acceptance. First, cluster-style maintenance over many nodes is typically performed a few nodes at a time, mak-ing maintenance slow and often impractical. Second, cluster-style maintenance does not work on single-node systems, despite the fact that their unavailability during maintenance can be painful for organizations. In this paper, we propose a novel technique for online maintenance that uses virtual machines to provide maintenance on single nodes, allowing parallel maintenance over multiple nodes, and online maintenance for standalone servers. We present the Microvisor, our prototype virtual machine system that is custom tailored to the needs of online maintenance. Unlike general purpose virtual machine environments that induce continual 10-20% over-head, the Microvisor virtualizes the hardware only during periods of active maintenance, letting the guest OS run at full speed most of the time. Unlike past attempts at virtual machine optimization, we do not compromise OS transparency. We instead give up generality and tailor our virtual machine system to the minimum needs of online maintenance, eschewing features, such as I/O and memory virtualization, that it does not strictly require. The result is a very thin virtual machine system that induces only 5.6% CPU overhead when virtualizing the hardware, and zero CPU overhead when devirtualized . Using the Microvisor, we demonstrate an online OS upgrade on a live, single-node web server, reducing downtime from one hour to less than one minute.