Scalable deep learning remains an onerous challenge, as it is constrained by many factors, including those related to load imbalance. For many deep-learning software systems, multiple data-processing components-including neural network training, graph scheduling, input pipeline, and gradient synchronization-execute simultaneously and asynchronously. Such execution can cause the various data-processing components to contend with one another for the hardware resources, leading to severe load imbalance and, in turn, degraded scalability. In this paper, we present an in-depth analysis of state-of-the-art deep-learning software, TensorFlow and Horovod, to understand their scalability limitations. Based on this analysis, we propose four novel solutions that minimize resource contention and improve deep-learning performance by up to 35% for training various neural networks on 24,576 GPUs of the Summit supercomputer at Oak Ridge National Laboratory.

Productivity from day one on supercomputers that leverage new technologies requires significant preparation. An institution that procures a novel system architecture often lacks sufficient institutional knowledge and skills to prepare for it. Thus, the "Center of Excellence" (CoE) concept has emerged to prepare for systems such as Summit and Sierra, currently the top two systems in the Top 500. This paper documents CoE experiences that prepared a workload of diverse applications and math libraries for a heterogeneous system. We describe our approach to this preparation, including our management and execution strategies, and detail our experiences with and reasons for using different programming approaches. Our early science and performance results show that the project enabled significant early seismic science with up to a l4X throughput increase over Cori. In addition to our successes, we discuss our challenges and failures so others may benefit from our experience.

A graph application written using a distributed graph processing framework can perform over an order of magnitude slower than its high-performance, native counterpart. This issue stems from the aim, common to most graph frameworks, of restricting the scope of application development to specific graph constructs, such as, for example, vertex or edge programs.

Developments in machine learning and graph analytics have seen these fields establish themselves as pervasive in a wide range of applications. Non-volatile memory (NVRAM) offers higher capacity and information retainment in case of power loss, therefore it is expected to be adopted for such applications. However, the asymmetric access latencies of NVRAM greatly degrade performance. The focus of this paper is to reduce the effect of memory access latency on emerging machine learning and graph workloads. The proposed mechanism uses software tagging of application data structures so as to control on-chip cache evictions based on data type and reuse patterns in an NVRAM based multicore system. Learner models are developed that are capable of predicting cache allocations for a variety of machine learning and graph applications. The optimized learning model yields an average performance benefit of 21% compared to a system that does not optimize for the write latency challenges in NVRAM.

Sparse Matrix-Vector multiplication (SpMV) is a fundamental kernel, used by a large class of numerical algorithms. Emerging big-data and machine learning applications are propelling a renewed interest in SpMV algorithms that can tackle massive amount of unstructured data---rapidly approaching the TeraByte range---with predictable, high performance. In this paper we describe a new methodology to design SpMV algorithms for shared memory multiprocessors (SMPs) that organizes the original SpMV algorithm into two distinct phases. In the first phase we build a scaled matrix, that is reduced in the second phase, providing numerous opportunities to exploit memory locality. Using this methodology, we have designed two algorithms. Our experiments on irregular big-data matrices (an order of magnitude larger than the current state of the art) show a quasi-optimal scaling on a large-scale POWER8 SMP system, with an average performance speedup of 3.8x, when compared to an equally optimized version of the CSR algorithm. In terms of absolute performance, with our implementation, the POWER8 SMP system is comparable to a 256-node cluster. In terms of size, it can process matrices with up to 68 billion edges, an order of magnitude larger than state-of-the-art clusters.

The problem of counting occurrences of query graphs in a large data graph, known as subgraph counting, is fundamental to several domains such as genomics and social network analysis. Many important special cases (e.g. triangle counting) have received significant attention. Color coding is a very general and powerful algorithmic technique for subgraph counting. Color coding has been shown to be effective in several applications, but scalable implementations are only known for the special case of tree queries (i.e. queries of treewidth one). In this paper we present the first efficient distributed implementation for color coding that goes beyond tree queries: ouralgorithm applies to any query graph of treewidth 2. Since tree queries can be solved in time linear in the size of the data graph, our contribution is the first step into the realm of color codingfor queries that require superlinear worst case running time. This superlinear complexity leads to significant load balancing problems on graphs with heavy tailed degree distributions. Our algorithm works around high degree nodes in the data graph, and achieves very good runtime and scalability on a diverse collection of data and query graph pairs. We also provide a theoretical analysis of our algorithmic techniques, exhibiting asymptotic improvements in runtime on random graphs with power law degree distributions, a popular model for real world graphs.

In this paper we evaluate the performance of a large-scale POWER8 symmetric multiprocessor (SMP) system with eight processors. We focus our attention on cache and memory subsystems, analyzing the characteristics that have a direct impact on high-performance computing and analytics applications. We provide insight into the relevant characteristics of the POWER8 processor using a set of micro-benchmarks. We also analyze the POWER8 SMP at the system level using the well-known roofline model. Using the knowledge gained from these micro-benchmarks, we optimize three applications and use them to assess the capabilities of the POWER8 system. The results show that the POWER8 - based SMP system is capable of delivering high performance for a wide range of applications and kernels.

In this paper we present and evaluate a parallel community detection algorithm derived from the state-of-the-art Louvain modularity maximization method. Our algorithm adopts a novel graph mapping and data representation, and relies on can efficient communication runtime, specifically designed for fine-grained applications executed on large-scale supercomputers. We have been able to parallelize graphs with up to 138 billion edges on 8, 192 Blue Gene/Q nodes and 1, 024 P7-IH nodes. Leveraging the convergence properties of our algorithm and the efficient implementation, we can analyze communities of large scale graphs in just a few seconds. To the best of our knowledge, this is the first parallel implementation of the Louvain algorithm that scales to these large data and processor configurations.

Graph analytics are arguably one of the most demanding workloads for high-performance systems and interconnection networks. Graph applications often display all-to-all, fine-grained, high-rate communication patterns that expose the limits of the network protocol stacks. Load and communication imbalance generate hard-to-predict network hot-spots, and may require computational steering due to unpredictable data distributions. In this paper we present a lightweight communication library, implemented "on the metal" of BlueGene/Q and POWER7 IH that we have used to support large-scale graph algorithms up to 96K processing nodes and 6 million threads. With this library we have explored several optimization techniques, including overlapped communication, non-blocking collectives, message aggregation, and computation in the network for special collective communication patterns, such as parallel prefix. The experimental results show significant performance improvements, ranging from 5X to 10X, when compared to equally optimized MPI implementations.

MapReduce is a popular parallel processing framework for large-scale data analytics. To keep up with the increasing volume of datasets, it requires efficient I/O capability from the underlying computer systems to process and analyze data in two phases (mapping and reducing). Between these phases, MapReduce requires a shuffling phase to globally exchange the intermediate data generated by the mapping phase. We reveal that data shuffling, by physically moving segments of intermediate data across disks, causes significant I/O contention and compounds the I/O problem. In this paper, we propose a novel virtual shuffling strategy to enable efficient data movement and reduce I/O for MapReduce shuffling, thereby reducing power consumption and conserving energy. Virtual shuffling is realized through a combination of three techniques including a three-level segment table, near-demand merging, and dynamic and balanced merging subtrees. Our experimental results show that virtual shuffling significantly speeds up data movement in MapReduce and achieves faster job execution. Particularly, its reduction in disk I/O accesses results in as much as 12% savings in power consumption for MapReduce programs.

Hadoop is a popular open source implementation of the MapReduce programming model for cloud computing. However, it faces a number of issues to achieve the best performance from the underlying systems. These include a serialization barrier that delays the reduce phase, repetitive merges, and disk accesses, and the lack of portability to different interconnects. To keep up with the increasing volume of data sets, Hadoop also requires efficient I/O capability from the underlying computer systems to process and analyze data. We describe Hadoop-A, an acceleration framework that optimizes Hadoop with plug-in components for fast data movement, overcoming the existing limitations. A novel network-levitated merge algorithm is introduced to merge data without repetition and disk access. In addition, a full pipeline is designed to overlap the shuffle, merge, and reduce phases. Our experimental results show that Hadoop-A significantly speeds up data movement in MapReduce and doubles the throughput of Hadoop. In addition, Hadoop-A significantly reduces disk accesses caused by intermediate data.

One-sided communication is important to enable asynchronous communication and data movement for Global Address Space (GAS) programming models. Such communication is typically realized through direct messages between initiator and target processes. For peta scale systems with 10,000s of nodes and 100,000s of cores, these direct messages require dedicated communication buffers and/or channels, which can lead to significant scalability challenges for GAS programming models. In this paper, we describe a network-friendly communication model, multinode cooperation, to enable indirect one-sided communication. Compute nodes work together to handle one-side requests through (1) request forwarding in which one node can intercept a request and forward it to a target node, and (2) request aggregation in which one node can aggregate many requests to a target node. We have implemented multinode cooperation for a popular GAS runtime library, Aggregate Remote Memory Copy Interface (ARMCI). Our experimental results on a large scale Cray XT5 system demonstrate that multinode cooperations able to greatly increase memory scalability by reducing communication buffers required on each node. In addition, multinode cooperation improves the resiliency of GAS runtime system to network contention. Furthermore, multinode cooperation can benefit the performance of scientific applications. In one case, it reduces the total execution time of an NWChem application by 52%.

Global Address Space (GAS) programming models enable a convenient, shared-memory style addressing model, and support completely asynchronous data movement. Their underlying runtime systems face critical challenges in (1) scalably managing resources (such as memory for communication buffers), and (2) gracefully handling unpredictable communication patterns and any associated contention. In this research, we investigate these challenges for a popular GAS runtime library, Aggregate Remote Memory Copy Interface (ARMCI) on, large-scale Cray XT5 systems. We represent the management of communication resources as directed graphs, and propose two new scalable virtual topologies, Meshed Fully Connected Graphs (MFCG) and Cubic Fully Connected Graphs (CFCG), for scalable resource management and contention attenuation. To ensure deadlock-free communication in these multi-dimensional topologies, we design and develop Lowest Dimension First (LDF) forwarding to support fully- or partially-populated MFCG and CFCG on any number of nodes. We have extensively evaluated the benefits of these virtual topologies on the petascale Jaguar Cray XT5 system at Oak Ridge National Laboratory. Our experimental results demonstrate MFCG as the most suitable virtual topology because of its benefits in resource management, contention mitigation, and the resulting benefit to scientific applications.

Hadoop is a popular open-source implementation of the MapReduce programming model for cloud computing. However, it faces a number of issues to achieve the best performance from the underlying system. These include a serialization barrier that delays the reduce phase, repetitive merges and disk access, and lack of capability to leverage latest high speed interconnects. We describe Hadoop-A, an acceleration framework that optimizes Hadoop with plugin components implemented in C++ for fast data movement, overcoming its existing limitations. A novel network-levitated merge algorithm is introduced to merge data without repetition and disk access. In addition, a full pipeline is designed to overlap the shuffle, merge and reduce phases. Our experimental results show that Hadoop-A doubles the data processing throughput of Hadoop, and reduces CPU utilization by more than 36%.