Search

Real-time supercomputing performance analysis is a critical aspect of evaluating and optimizing computational systems in a dynamic user environment. The operation of supercomputers produce vast quantities of analytic data from multiple sources and of varying types so compiling this data in an efficient matter is critical to the process. MIT Lincoln Laboratory Supercomputing Center has been utilizing the Unity 3D game engine to create a Digital Twin of our supercomputing systems for several years to perform system monitoring. Unity offers robust visualization capabilities making it ideal for creating a sophisticated representation of the computational processes. As we scale the systems to include a diversity of resources such as accelerators and the addition of more users, we need to implement new analysis tools for the monitoring system. The workloads in research continuously change, as does the capability of Unity, and this allows us to adapt our monitoring tools to scale and incorporate features enabling efficient replay of system wide events, user isolation, and machine level granularity. Our system fully takes advantage of the modern capabilities of the Unity Engine in a way that intuitively represents the real time workload performed on a supercomputer. It allows HPC system engineers to quickly diagnose usage related errors with its responsive user interface which scales efficiently with large data sets.

Hypersparse traffic matrices constructed from network packet source and destination addresses is a powerful tool for gaining insights into network traffic. SuiteSparse: GraphBLAS, an open source package or building, manipulating, and analyzing large hypersparse matrices, is one approach to constructing these traffic matrices. Suricata is a widely used open source network intrusion detection software package. This work demonstrates how Suricata network flow records can be used to efficiently construct hypersparse matrices using GraphBLAS.

HPC systems used for research run a wide variety of software and workflows. This software is often written or modified by users to meet the needs of their research projects, and rarely is built with security in mind. In this paper we explore several of the key techniques that MIT Lincoln Laboratory Supercomputing Center has deployed on its systems to manage the security implications of these workflows by providing enforced separation for processes, filesystem access, network traffic, and accelerators to make every user feel like they are running on a personal HPC.

One of the more complex tasks for researchers using HPC systems is performance monitoring and tuning of their applications. Developing a practice of continuous performance improvement, both for speed-up and efficient use of resources is essential to the long term success of both the HPC practitioner and the research project. Profiling tools provide a nice view of the performance of an application but often have a steep learning curve and rarely provide an easy to interpret view of resource utilization. Lower level tools such as top and htop provide a view of resource utilization for those familiar and comfortable with Linux but a barrier for newer HPC practitioners. To expand the existing profiling and job monitoring options, the MIT Lincoln Laboratory Supercomputing Center created LLoad, a tool that captures a snapshot of the resources being used by a job on a per user basis. LLload is a tool built from standard HPC tools that provides an easy way for a researcher to track resource usage of active jobs. We explain how the tool was designed and implemented and provide insight into how it is used to aid new researchers in developing their performance monitoring skills as well as guide researchers in their resource requests.

As research and deployment of AI grows, the computational burden to support and sustain its progress inevitably does too. To train or fine-tune state-of-the-art models in NLP, computer vision, etc., some form of AI hardware acceleration is virtually a requirement. Recent large language models require considerable resources to train and deploy, resulting in significant energy usage, potential carbon emissions, and massive demand for GPUs and other hardware accelerators. However, this surge carries large implications for energy sustainability at the HPC/datacenter level. In this paper, we study the aggregate effect of power-capping GPUs on GPU temperature and power draw at a research supercomputing center. With the right amount of power-capping, we show significant decreases in both temperature and power draw, reducing power consumption and potentially improving hardware life-span with minimal impact on job performance. While power-capping reduces power draw by design, the aggregate system-wide effect on overall energy consumption is less clear; for instance, if users notice job performance degradation from GPU power-caps, they may request additional GPU-jobs to compensate, negating any energy savings or even worsening energy consumption. To our knowledge, our work is the first to conduct and make available a detailed analysis of the effects of GPU power-capping at the supercomputing scale. We hope our work will inspire HPCs/datacenters to further explore, evaluate, and communicate the impact of power-capping AI hardware accelerators for more sustainable AI.

Low-power small form factor data processing units (DPUs) enable offloading and acceleration of a broad range of networking and security services. DPUs have accelerated the transition to programmable networking by enabling the replacement of FPGAs/ASICs in a wide range of network oriented devices. The GraphBLAS sparse matrix graph open standard math library is well-suited for constructing anonymized hypersparse traffic matrices of network traffic which can enable a wide range of network analytics. This paper measures the performance of the GraphBLAS on an ARM based NVIDIA DPU (BlueField 2) and, to the best of our knowledge, represents the first reported GraphBLAS results on a DPU and/or ARM based system. Anonymized hypersparse traffic matrices were constructed at a rate of over 18 million packets per second.

This paper is an update of the survey of AI accelerators and processors from past four years, which is now called the Lincoln AI Computing Survey - LAICS (pronounced "lace"). As in past years, this paper collects and summarizes the current commercial accelerators that have been publicly announced with peak performance and peak power consumption numbers. The performance and power values are plotted on a scatter graph, and a number of dimensions and observations from the trends on this plot are again discussed and analyzed. Market segments are highlighted on the scatter plot, and zoomed plots of each segment are also included. Finally, a brief description of each of the new accelerators that have been added in the survey this year is included., Comment: 7 pages, 6 figures, 2023 IEEE High Performance Extreme Computing (HPEC) conference, September 2023

Large language models (LLMs) have exploded in popularity due to their new generative capabilities that go far beyond prior state-of-the-art. These technologies are increasingly being leveraged in various domains such as law, finance, and medicine. However, these models carry significant computational challenges, especially the compute and energy costs required for inference. Inference energy costs already receive less attention than the energy costs of training LLMs -- despite how often these large models are called on to conduct inference in reality (e.g., ChatGPT). As these state-of-the-art LLMs see increasing usage and deployment in various domains, a better understanding of their resource utilization is crucial for cost-savings, scaling performance, efficient hardware usage, and optimal inference strategies. In this paper, we describe experiments conducted to study the computational and energy utilization of inference with LLMs. We benchmark and conduct a preliminary analysis of the inference performance and inference energy costs of different sizes of LLaMA -- a recent state-of-the-art LLM -- developed by Meta AI on two generations of popular GPUs (NVIDIA V100 \& A100) and two datasets (Alpaca and GSM8K) to reflect the diverse set of tasks/benchmarks for LLMs in research and practice. We present the results of multi-node, multi-GPU inference using model sharding across up to 32 GPUs. To our knowledge, our work is the one of the first to study LLM inference performance from the perspective of computational and energy resources at this scale.

pPython seeks to provide a parallel capability that provides good speed-up without sacrificing the ease of programming in Python by implementing partitioned global array semantics (PGAS) on top of a simple file-based messaging library (PythonMPI) in pure Python. pPython follows a SPMD (single program multiple data) model of computation. pPython runs on a single-node (e.g., a laptop) running Windows, Linux, or MacOS operating systems or on any combination of heterogeneous systems that support Python, including on a cluster through a Slurm scheduler interface so that pPython can be executed in a massively parallel computing environment. It is interesting to see what performance pPython can achieve compared to the traditional socket-based MPI communication because of its unique file-based messaging implementation. In this paper, we present the point-to-point and collective communication performances of pPython and compare them with those obtained by using mpi4py with OpenMPI. For large messages, pPython demonstrates comparable performance as compared to mpi4py., Comment: arXiv admin note: substantial text overlap with arXiv:2208.14908

As research and practice in artificial intelligence (A.I.) grow in leaps and bounds, the resources necessary to sustain and support their operations also grow at an increasing pace. While innovations and applications from A.I. have brought significant advances, from applications to vision and natural language to improvements to fields like medical imaging and materials engineering, their costs should not be neglected. As we embrace a world with ever-increasing amounts of data as well as research and development of A.I. applications, we are sure to face an ever-mounting energy footprint to sustain these computational budgets, data storage needs, and more. But, is this sustainable and, more importantly, what kind of setting is best positioned to nurture such sustainable A.I. in both research and practice? In this paper, we outline our outlook for Green A.I. -- a more sustainable, energy-efficient and energy-aware ecosystem for developing A.I. across the research, computing, and practitioner communities alike -- and the steps required to arrive there. We present a bird's eye view of various areas for potential changes and improvements from the ground floor of AI's operational and hardware optimizations for datacenters/HPCs to the current incentive structures in the world of A.I. research and practice, and more. We hope these points will spur further discussion, and action, on some of these issues and their potential solutions., Comment: 8 pages, published in 2022 IEEE International Parallel and Distributed Processing Symposium Workshops (IPDPSW)

Internet analysis is a major challenge due to the volume and rate of network traffic. In lieu of analyzing traffic as raw packets, network analysts often rely on compressed network flows (netflows) that contain the start time, stop time, source, destination, and number of packets in each direction. However, many traffic analyses benefit from temporal aggregation of multiple simultaneous netflows, which can be computationally challenging. To alleviate this concern, a novel netflow compression and resampling method has been developed leveraging GraphBLAS hyperspace traffic matrices that preserve anonymization while enabling subrange analysis. Standard multitemporal spatial analyses are then performed on each subrange to generate detailed statistical aggregates of the source packets, source fan-out, unique links, destination fan-in, and destination packets of each subrange which can then be used for background modeling and anomaly detection. A simple file format based on GraphBLAS sparse matrices is developed for storing these statistical aggregates. This method is scale tested on the MIT SuperCloud using a 50 trillion packet netflow corpus from several hundred sites collected over several months. The resulting compression achieved is significant (<0.1 bit per packet) enabling extremely large netflow analyses to be stored and transported. The single node parallel performance is analyzed in terms of both processors and threads showing that a single node can perform hundreds of simultaneous analyses at over a million packets/sec (roughly equivalent to a 10 Gigabit link)., Comment: arXiv admin note: text overlap with arXiv:2203.13934, arXiv:2108.06653, arXiv:2008.00307

In this paper we address the application of pre-processing techniques to multi-channel time series data with varying lengths, which we refer to as the alignment problem, for downstream machine learning. The misalignment of multi-channel time series data may occur for a variety of reasons, such as missing data, varying sampling rates, or inconsistent collection times. We consider multi-channel time series data collected from the MIT SuperCloud High Performance Computing (HPC) center, where different job start times and varying run times of HPC jobs result in misaligned data. This misalignment makes it challenging to build AI/ML approaches for tasks such as compute workload classification. Building on previous supervised classification work with the MIT SuperCloud Dataset, we address the alignment problem via three broad, low overhead approaches: sampling a fixed subset from a full time series, performing summary statistics on a full time series, and sampling a subset of coefficients from time series mapped to the frequency domain. Our best performing models achieve a classification accuracy greater than 95%, outperforming previous approaches to multi-channel time series classification with the MIT SuperCloud Dataset by 5%. These results indicate our low overhead approaches to solving the alignment problem, in conjunction with standard machine learning techniques, are able to achieve high levels of classification accuracy, and serve as a baseline for future approaches to addressing the alignment problem, such as kernel methods.

Python has become a standard scientific computing language with fast-growing support of machine learning and data analysis modules, as well as an increasing usage of big data. The Dynamic Distributed Dimensional Data Model (D4M) offers a highly composable, unified data model with strong performance built to handle big data fast and efficiently. In this work we present an implementation of D4M in Python. $D4M.py$ implements all foundational functionality of D4M and includes Accumulo and SQL database support via Graphulo. We describe the mathematical background and motivation, an explanation of the approaches made for its fundamental functions and building blocks, and performance results which compare $D4M.py$'s performance to D4M-MATLAB and D4M.jl., Comment: 8 pages, 7 figures, accepted to HPEC 2022

pPython seeks to provide a parallel capability that provides good speed-up without sacrificing the ease of programming in Python by implementing partitioned global array semantics (PGAS) on top of a simple file-based messaging library (PythonMPI) in pure Python. The core data structure in pPython is a distributed numerical array whose distribution onto multiple processors is specified with a map construct. Communication operations between distributed arrays are abstracted away from the user and pPython transparently supports redistribution between any block-cyclic-overlapped distributions in up to four dimensions. pPython follows a SPMD (single program multiple data) model of computation. pPython runs on any combination of heterogeneous systems that support Python, including Windows, Linux, and MacOS operating systems. In addition to running transparently on single-node (e.g., a laptop), pPython provides a scheduler interface, so that pPython can be executed in a massively parallel computing environment. The initial implementation uses the Slurm scheduler. Performance of pPython on the HPC Challenge benchmark suite demonstrates both ease of programming and scalability., Comment: arXiv admin note: substantial text overlap with arXiv:astro-ph/0606464

High-Performance Computing (HPC) centers and cloud providers support an increasingly diverse set of applications on heterogenous hardware. As Artificial Intelligence (AI) and Machine Learning (ML) workloads have become an increasingly larger share of the compute workloads, new approaches to optimized resource usage, allocation, and deployment of new AI frameworks are needed. By identifying compute workloads and their utilization characteristics, HPC systems may be able to better match available resources with the application demand. By leveraging datacenter instrumentation, it may be possible to develop AI-based approaches that can identify workloads and provide feedback to researchers and datacenter operators for improving operational efficiency. To enable this research, we released the MIT Supercloud Dataset, which provides detailed monitoring logs from the MIT Supercloud cluster. This dataset includes CPU and GPU usage by jobs, memory usage, and file system logs. In this paper, we present a workload classification challenge based on this dataset. We introduce a labelled dataset that can be used to develop new approaches to workload classification and present initial results based on existing approaches. The goal of this challenge is to foster algorithmic innovations in the analysis of compute workloads that can achieve higher accuracy than existing methods. Data and code will be made publicly available via the Datacenter Challenge website : https://dcc.mit.edu., Comment: Accepted at IPDPS ADOPT'22

Long range detection is a cornerstone of defense in many operating domains (land, sea, undersea, air, space, ..,). In the cyber domain, long range detection requires the analysis of significant network traffic from a variety of observatories and outposts. Construction of anonymized hypersparse traffic matrices on edge network devices can be a key enabler by providing significant data compression in a rapidly analyzable format that protects privacy. GraphBLAS is ideally suited for both constructing and analyzing anonymized hypersparse traffic matrices. The performance of GraphBLAS on an Accolade Technologies edge network device is demonstrated on a near worse case traffic scenario using a continuous stream of CAIDA Telescope darknet packets. The performance for varying numbers of traffic buffers, threads, and processor cores is explored. Anonymized hypersparse traffic matrices can be constructed at a rate of over 50,000,000 packets per second; exceeding a typical 400 Gigabit network link. This performance demonstrates that anonymized hypersparse traffic matrices are readily computable on edge network devices with minimal compute resources and can be a viable data product for such devices., Comment: Accepted to IEEE HPEC, Outstanding Paper Award, 8 pages, 8 figures, 1 table, 70 references. arXiv admin note: text overlap with arXiv:2108.06653, arXiv:2008.00307, arXiv:2203.10230

The Internet has become a critical component of modern civilization requiring scientific exploration akin to endeavors to understand the land, sea, air, and space environments. Understanding the baseline statistical distributions of traffic are essential to the scientific understanding of the Internet. Correlating data from different Internet observatories and outposts can be a useful tool for gaining insights into these distributions. This work compares observed sources from the largest Internet telescope (the CAIDA darknet telescope) with those from a commercial outpost (the GreyNoise honeyfarm). Neither of these locations actively emit Internet traffic and provide distinct observations of unsolicited Internet traffic (primarily botnets and scanners). Newly developed GraphBLAS hyperspace matrices and D4M associative array technologies enable the efficient analysis of these data on significant scales. The CAIDA sources are well approximated by a Zipf-Mandelbrot distribution. Over a 6-month period 70\% of the brightest (highest frequency) sources in the CAIDA telescope are consistently detected by coeval observations in the GreyNoise honeyfarm. This overlap drops as the sources dim (reduce frequency) and as the time difference between the observations grows. The probability of seeing a CAIDA source is proportional to the logarithm of the brightness. The temporal correlations are well described by a modified Cauchy distribution. These observations are consistent with a correlated high frequency beam of sources that drifts on a time scale of a month., Comment: 8 pages, 8 figures, 2 tables, 59 references; accepted to GrAPL 2022. arXiv admin note: substantial text overlap with arXiv:2108.06653

Deep learning (DL) workflows demand an ever-increasing budget of compute and energy in order to achieve outsized gains. Neural architecture searches, hyperparameter sweeps, and rapid prototyping consume immense resources that can prevent resource-constrained researchers from experimenting with large models and carry considerable environmental impact. As such, it becomes essential to understand how different deep neural networks (DNNs) and training leverage increasing compute and energy resources -- especially specialized computationally-intensive models across different domains and applications. In this paper, we conduct over 3,400 experiments training an array of deep networks representing various domains/tasks -- natural language processing, computer vision, and chemistry -- on up to 424 graphics processing units (GPUs). During training, our experiments systematically vary compute resource characteristics and energy-saving mechanisms such as power utilization and GPU clock rate limits to capture and illustrate the different trade-offs and scaling behaviors each representative model exhibits under various resource and energy-constrained regimes. We fit power law models that describe how training time scales with available compute resources and energy constraints. We anticipate that these findings will help inform and guide high-performance computing providers in optimizing resource utilization, by selectively reducing energy consumption for different deep learning tasks/workflows with minimal impact on training., Comment: 14 pages, 17 figures

The Internet is transforming our society, necessitating a quantitative understanding of Internet traffic. Our team collects and curates the largest publicly available Internet traffic data sets. An analysis of 50 billion packets using 10,000 processors in the MIT SuperCloud reveals a new phenomenon: the importance of otherwise unseen leaf nodes and isolated links in Internet traffic. Our analysis further shows that a two-parameter modified Zipf-Mandelbrot distribution accurately describes a wide variety of source/destination statistics on moving sample windows ranging from 100{,}000 to 100{,}000{,}000 packets over collections that span years and continents. The measured model parameters distinguish different network streams, and the model leaf parameter strongly correlates with the fraction of the traffic in different underlying network topologies., Comment: 53 pages, 27 figures, 8 tables, 121 references. Portions of this work originally appeared as arXiv:1904.04396v1 which has been split for publication in the book "Massive Graph Analytics" (edited by David Bader)

Over the past several years, new machine learning accelerators were being announced and released every month for a variety of applications from speech recognition, video object detection, assisted driving, and many data center applications. This paper updates the survey of AI accelerators and processors from past two years. This paper collects and summarizes the current commercial accelerators that have been publicly announced with peak performance and power consumption numbers. The performance and power values are plotted on a scatter graph, and a number of dimensions and observations from the trends on this plot are again discussed and analyzed. This year, we also compile a list of benchmarking performance results and compute the computational efficiency with respect to peak performance., Comment: 9 pages, 2 figures, IEEE High Performance Extreme Computing Conference 2021

Supercomputers are complex systems producing vast quantities of performance data from multiple sources and of varying types. Performance data from each of the thousands of nodes in a supercomputer tracks multiple forms of storage, memory, networks, processors, and accelerators. Optimization of application performance is critical for cost effective usage of a supercomputer and requires efficient methods for effectively viewing performance data. The combination of supercomputing analytics and 3D gaming visualization enables real-time processing and visual data display of massive amounts of information that humans can process quickly with little training. Our system fully utilizes the capabilities of modern 3D gaming environments to create novel representations of computing hardware which intuitively represent the physical attributes of the supercomputer while displaying real-time alerts and component utilization. This system allows operators to quickly assess how the supercomputer is being used, gives users visibility into the resources they are consuming, and provides instructors new ways to interactively teach the computing architecture concepts necessary for efficient computing

First responders and other forward deployed essential workers can benefit from advanced analytics. Limited network access and software security requirements prevent the usage of standard cloud based microservice analytic platforms that are typically used in industry. One solution is to precompute a wide range of analytics as files that can be used with standard preinstalled software that does not require network access or additional software and can run on a wide range of legacy hardware. In response to the COVID-19 pandemic, this approach was tested for providing geo-spatial census data to allow quick analysis of demographic data for better responding to emergencies. These data were processed using the MIT SuperCloud to create several thousand Google Earth and Microsoft Excel files representative of many advanced analytics. The fast mapping of census data using Google Earth and Microsoft Excel has the potential to give emergency responders a powerful tool to improve emergency preparedness. Our approach displays relevant census data (total population, population under 15, population over 65, median age) per census block, sorted by county, through a Microsoft Excel spreadsheet (xlsx file) and Google Earth map (kml file). The spreadsheet interface includes features that allow users to convert between different longitude and latitude coordinate units. For the Google Earth files, a variety of absolute and relative colors maps of population density have been explored to provide an intuitive and meaningful interface. Using several hundred cores on the MIT SuperCloud, new analytics can be generated in a few minutes., Comment: 5 pages, 11 figures, 17 references, accepted to IEEE HPEC 2021

AI algorithms that identify maneuvers from trajectory data could play an important role in improving flight safety and pilot training. AI challenges allow diverse teams to work together to solve hard problems and are an effective tool for developing AI solutions. AI challenges are also a key driver of AI computational requirements. The Maneuver Identification Challenge hosted at maneuver-id.mit.edu provides thousands of trajectories collected from pilots practicing in flight simulators, descriptions of maneuvers, and examples of these maneuvers performed by experienced pilots. Each trajectory consists of positions, velocities, and aircraft orientations normalized to a common coordinate system. Construction of the data set required significant data architecture to transform flight simulator logs into AI ready data, which included using a supercomputer for deduplication and data conditioning. There are three proposed challenges. The first challenge is separating physically plausible (good) trajectories from unfeasible (bad) trajectories. Human labeled good and bad trajectories are provided to aid in this task. Subsequent challenges are to label trajectories with their intended maneuvers and to assess the quality of those maneuvers., Comment: 7 pages, 8 figures, 1 table, 33 references, accepted to IEEE HPEC 2021

Diverse workloads such as interactive supercomputing, big data analysis, and large-scale AI algorithm development, requires a high-performance scheduler. This paper presents a novel node-based scheduling approach for large scale simulations of short running jobs on MIT SuperCloud systems, that allows the resources to be fully utilized for both long running batch jobs while simultaneously providing fast launch and release of large-scale short running jobs. The node-based scheduling approach has demonstrated up to 100 times faster scheduler performance that other state-of-the-art systems., Comment: IEEE HPEC 2021

The Internet has never been more important to our society, and understanding the behavior of the Internet is essential. The Center for Applied Internet Data Analysis (CAIDA) Telescope observes a continuous stream of packets from an unsolicited darkspace representing 1/256 of the Internet. During 2019 and 2020 over 40,000,000,000,000 unique packets were collected representing the largest ever assembled public corpus of Internet traffic. Using the combined resources of the Supercomputing Centers at UC San Diego, Lawrence Berkeley National Laboratory, and MIT, the spatial temporal structure of anonymized source-destination pairs from the CAIDA Telescope data has been analyzed with GraphBLAS hierarchical hypersparse matrices. These analyses provide unique insight on this unsolicited Internet darkspace traffic with the discovery of many previously unseen scaling relations. The data show a significant sustained increase in unsolicited traffic corresponding to the start of the COVID19 pandemic, but relatively little change in the underlying scaling relations associated with unique sources, source fan-outs, unique links, destination fan-ins, and unique destinations. This work provides a demonstration of the practical feasibility and benefit of the safe collection and analysis of significant quantities of anonymized Internet traffic., Comment: 8 pages, 9 figures, 2 tables, 43 references, accepted to IEEE HPEC 2021. arXiv admin note: substantial text overlap with arXiv:2008.00307

Hypersparse matrices are a powerful enabler for a variety of network, health, finance, and social applications. Hierarchical hypersparse GraphBLAS matrices enable rapid streaming updates while preserving algebraic analytic power and convenience. In many contexts, the rate of these updates sets the bounds on performance. This paper explores hierarchical hypersparse update performance on a variety of hardware with identical software configurations. The high-level language bindings of the GraphBLAS readily enable performance experiments on simultaneous diverse hardware. The best single process performance measured was 4,000,000 updates per second. The best single node performance measured was 170,000,000 updates per second. The hardware used spans nearly a decade and allows a direct comparison of hardware improvements for this computation over this time range; showing a 2x increase in single-core performance, a 3x increase in single process performance, and a 5x increase in single node performance. Running on nearly 2,000 MIT SuperCloud nodes simultaneously achieved a sustained update rate of over 200,000,000,000 updates per second. Hierarchical hypersparse GraphBLAS allows the MIT SuperCloud to analyze extremely large streaming network data sets., Comment: 6 pages, 5 figures, 32 references, accepted to IEEE HPEC 2021. arXiv admin note: text overlap with arXiv:2001.06935

Artificial intelligence (AI) and Machine learning (ML) workloads are an increasingly larger share of the compute workloads in traditional High-Performance Computing (HPC) centers and commercial cloud systems. This has led to changes in deployment approaches of HPC clusters and the commercial cloud, as well as a new focus on approaches to optimized resource usage, allocations and deployment of new AI frame- works, and capabilities such as Jupyter notebooks to enable rapid prototyping and deployment. With these changes, there is a need to better understand cluster/datacenter operations with the goal of developing improved scheduling policies, identifying inefficiencies in resource utilization, energy/power consumption, failure prediction, and identifying policy violations. In this paper we introduce the MIT Supercloud Dataset which aims to foster innovative AI/ML approaches to the analysis of large scale HPC and datacenter/cloud operations. We provide detailed monitoring logs from the MIT Supercloud system, which include CPU and GPU usage by jobs, memory usage, file system logs, and physical monitoring data. This paper discusses the details of the dataset, collection methodology, data availability, and discusses potential challenge problems being developed using this data. Datasets and future challenge announcements will be available via https://dcc.mit.edu.

New machine learning accelerators are being announced and released each month for a variety of applications from speech recognition, video object detection, assisted driving, and many data center applications. This paper updates the survey of of AI accelerators and processors from last year's IEEE-HPEC paper. This paper collects and summarizes the current accelerators that have been publicly announced with performance and power consumption numbers. The performance and power values are plotted on a scatter graph and a number of dimensions and observations from the trends on this plot are discussed and analyzed. For instance, there are interesting trends in the plot regarding power consumption, numerical precision, and inference versus training. This year, there are many more announced accelerators that are implemented with many more architectures and technologies from vector engines, dataflow engines, neuromorphic designs, flash-based analog memory processing, and photonic-based processing., Comment: 12 pages, 2 figures, IEEE-HPEC conference, Waltham, MA, September 21-25, 2020. arXiv admin note: text overlap with arXiv:1908.11348

Artificial Intelligence/Machine Learning applications require the training of complex models on large amounts of labelled data. The large computational requirements for training deep models have necessitated the development of new methods for faster training. One such approach is the data parallel approach, where the training data is distributed across multiple compute nodes. This approach is simple to implement and supported by most of the commonly used machine learning frameworks. The data parallel approach leverages MPI for communicating gradients across all nodes. In this paper, we examine the effects of using different physical hardware interconnects and network-related software primitives for enabling data distributed deep learning. We compare the effect of using GPUDirect and NCCL on Ethernet and OmniPath fabrics. Our results show that using Ethernet-based networking in shared HPC systems does not have a significant effect on the training times for commonly used deep neural network architectures or traditional HPC applications such as Computational Fluid Dynamics., Comment: Accepted for publication at IEEE HPEC 2020

Deep neural networks have shown great success in many diverse fields. The training of these networks can take significant amounts of time, compute and energy. As datasets get larger and models become more complex, the exploration of model architectures becomes prohibitive. In this paper we examine the compute, energy and time costs of training a UNet based deep neural network for the problem of predicting short term weather forecasts (called precipitation Nowcasting). By leveraging a combination of data distributed and mixed-precision training, we explore the design space for this problem. We also show that larger models with better performance come at a potentially incremental cost if appropriate optimizations are used. We show that it is possible to achieve a significant improvement in training time by leveraging mixed-precision training without sacrificing model performance. Additionally, we find that a 1549% increase in the number of trainable parameters for a network comes at a relatively smaller 63.22% increase in energy usage for a UNet with 4 encoding layers., Comment: Accepted for publication at IEEE HPEC 2020

Rapid launch of thousands of jobs is essential for effective interactive supercomputing, big data analysis, and AI algorithm development. Achieving thousands of launches per second has required hardware to be available to receive these jobs. This paper presents a novel preemptive approach to implement spot jobs on MIT SuperCloud systems allowing the resources to be fully utilized for both long running batch jobs while still providing fast launch for interactive jobs. The new approach separates the job preemption and scheduling operations and can achieve 100 times faster performance in the scheduling of a job with preemption when compared to using the standard scheduler-provided automatic preemption-based capability. The results demonstrate that the new approach can schedule interactive jobs preemptively at a performance comparable to when the required computing resources are idle and available. The spot job capability can be deployed without disrupting the interactive user experience while increasing the overall system utilization.

A Multigrid Full Approximation Storage algorithm for solving Deep Residual Networks is developed to enable neural network parallelized layer-wise training and concurrent computational kernel execution on GPUs. This work demonstrates a 10.2x speedup over traditional layer-wise model parallelism techniques using the same number of compute units., Comment: 7 pages, 6 figures, 27 citations. Accepted to 2020 IEEE High Performance Extreme Computing Conference - Outstanding Paper Award

Pandemic measures such as social distancing and contact tracing can be enhanced by rapidly integrating dynamic location data and demographic data. Projecting billions of longitude and latitude locations onto hundreds of thousands of highly irregular demographic census block polygons is computationally challenging in both research and deployment contexts. This paper describes two approaches labeled "simple" and "fast". The simple approach can be implemented in any scripting language (Matlab/Octave, Python, Julia, R) and is easily integrated and customized to a variety of research goals. This simple approach uses a novel combination of hierarchy, sparse bounding boxes, polygon crossing-number, vectorization, and parallel processing to achieve 100,000,000+ projections per second on 100 servers. The simple approach is compact, does not increase data storage requirements, and is applicable to any country or region. The fast approach exploits the thread, vector, and memory optimizations that are possible using a low-level language (C++) and achieves similar performance on a single server. This paper details these approaches with the goal of enabling the broader community to quickly integrate location and demographic data., Comment: 8 pages, 7 figures, 55 references; accepted to IEEE HPEC 2020

The rise of graph analytic systems has created a need for new ways to measure and compare the capabilities of graph processing systems. The MIT/Amazon/IEEE Graph Challenge has been developed to provide a well-defined community venue for stimulating research and highlighting innovations in graph analysis software, hardware, algorithms, and systems. GraphChallenge.org provides a wide range of pre-parsed graph data sets, graph generators, mathematically defined graph algorithms, example serial implementations in a variety of languages, and specific metrics for measuring performance. The triangle counting component of GraphChallenge.org tests the performance of graph processing systems to count all the triangles in a graph and exercises key graph operations found in many graph algorithms. In 2017, 2018, and 2019 many triangle counting submissions were received from a wide range of authors and organizations. This paper presents a performance analysis of the best performers of these submissions. These submissions show that their state-of-the-art triangle counting execution time, $T_{\rm tri}$, is a strong function of the number of edges in the graph, $N_e$, which improved significantly from 2017 ($T_{\rm tri} \approx (N_e/10^8)^{4/3}$) to 2018 ($T_{\rm tri} \approx N_e/10^9$) and remained comparable from 2018 to 2019. Graph Challenge provides a clear picture of current graph analysis systems and underscores the need for new innovations to achieve high performance on very large graphs., Comment: 10 pages, 8 figures, 121 references, to be submitted to IEEE HPEC 2020. This work reports new updated results on prior work reported in arXiv:1805.09675 & arXiv:1708.06866

The SuiteSparse GraphBLAS C-library implements high performance hypersparse matrices with bindings to a variety of languages (Python, Julia, and Matlab/Octave). GraphBLAS provides a lightweight in-memory database implementation of hypersparse matrices that are ideal for analyzing many types of network data, while providing rigorous mathematical guarantees, such as linearity. Streaming updates of hypersparse matrices put enormous pressure on the memory hierarchy. This work benchmarks an implementation of hierarchical hypersparse matrices that reduces memory pressure and dramatically increases the update rate into a hypersparse matrices. The parameters of hierarchical hypersparse matrices rely on controlling the number of entries in each level in the hierarchy before an update is cascaded. The parameters are easily tunable to achieve optimal performance for a variety of applications. Hierarchical hypersparse matrices achieve over 1,000,000 updates per second in a single instance. Scaling to 31,000 instances of hierarchical hypersparse matrices arrays on 1,100 server nodes on the MIT SuperCloud achieved a sustained update rate of 75,000,000,000 updates per second. This capability allows the MIT SuperCloud to analyze extremely large streaming network data sets., Comment: 4 pages, 4 figures, 28 references, accepted to IPDPS GrAPL 2020. arXiv admin note: substantial text overlap with arXiv:1907.04217

In this paper, we present a novel and new file-based communication architecture using the local filesystem for large scale parallelization. This new approach eliminates the issues with filesystem overload and resource contention when using the central filesystem for large parallel jobs. The new approach incurs additional overhead due to inter-node message file transfers when both the sending and receiving processes are not on the same node. However, even with this additional overhead cost, its benefits are far greater for the overall cluster operation in addition to the performance enhancement in message communications for large scale parallel jobs. For example, when running a 2048-process parallel job, it achieved about 34 times better performance with MPI_Bcast() when using the local filesystem. Furthermore, since the security for transferring message files is handled entirely by using the secure copy protocol (scp) and the file system permissions, no additional security measures or ports are required other than those that are typically required on an HPC system.

Federated authentication can drastically reduce the overhead of basic account maintenance while simultaneously improving overall system security. Integrating with the user's more frequently used account at their primary organization both provides a better experience to the end user and makes account compromise or changes in affiliation more likely to be noticed and acted upon. Additionally, with many organizations transitioning to multi-factor authentication for all account access, the ability to leverage external federated identity management systems provides the benefit of their efforts without the additional overhead of separately implementing a distinct multi-factor authentication process. This paper describes our experiences and the lessons we learned by enabling federated authentication with the U.S. Government PKI and InCommon Federation, scaling it up to the user base of a production HPC system, and the motivations behind those choices. We have received only positive feedback from our users.

The Dynamic Distributed Dimensional Data Model (D4M) library implements associative arrays in a variety of languages (Python, Julia, and Matlab/Octave) and provides a lightweight in-memory database implementation of hypersparse arrays that are ideal for analyzing many types of network data. D4M relies on associative arrays which combine properties of spreadsheets, databases, matrices, graphs, and networks, while providing rigorous mathematical guarantees, such as linearity. Streaming updates of D4M associative arrays put enormous pressure on the memory hierarchy. This work describes the design and performance optimization of an implementation of hierarchical associative arrays that reduces memory pressure and dramatically increases the update rate into an associative array. The parameters of hierarchical associative arrays rely on controlling the number of entries in each level in the hierarchy before an update is cascaded. The parameters are easily tunable to achieve optimal performance for a variety of applications. Hierarchical arrays achieve over 40,000 updates per second in a single instance. Scaling to 34,000 instances of hierarchical D4M associative arrays on 1,100 server nodes on the MIT SuperCloud achieved a sustained update rate of 1,900,000,000 updates per second. This capability allows the MIT SuperCloud to analyze extremely large streaming network data sets., Comment: 6 pages; 6 figures; accepted to IEEE High Performance Extreme Computing (HPEC) Conference 2019. arXiv admin note: text overlap with arXiv:1807.05308, arXiv:1902.00846

The Intel Xeon Phi manycore processor is designed to provide high performance matrix computations of the type often performed in data analysis. Common data analysis environments include Matlab, GNU Octave, Julia, Python, and R. Achieving optimal performance of matrix operations within data analysis environments requires tuning the Xeon Phi OpenMP settings, process pinning, and memory modes. This paper describes matrix multiplication performance results for Matlab and GNU Octave over a variety of combinations of process counts and OpenMP threads and Xeon Phi memory modes. These results indicate that using KMP_AFFINITY=granlarity=fine, taskset pinning, and all2all cache memory mode allows both Matlab and GNU Octave to achieve 66% of the practical peak performance for process counts ranging from 1 to 64 and OpenMP threads ranging from 1 to 64. These settings have resulted in generally improved performance across a range of applications and has enabled our Xeon Phi system to deliver significant results in a number of real-world applications., Comment: 6 pages, 5 figures, accepted in IEEE High Performance Extreme Computing (HPEC) conference 2019

For decades, the use of HPC systems was limited to those in the physical sciences who had mastered their domain in conjunction with a deep understanding of HPC architectures and algorithms. During these same decades, consumer computing device advances produced tablets and smartphones that allow millions of children to interactively develop and share code projects across the globe. As the HPC community faces the challenges associated with guiding researchers from disciplines using high productivity interactive tools to effective use of HPC systems, it seems appropriate to revisit the assumptions surrounding the necessary skills required for access to large computational systems. For over a decade, MIT Lincoln Laboratory has been supporting interactive, on-demand high performance computing by seamlessly integrating familiar high productivity tools to provide users with an increased number of design turns, rapid prototyping capability, and faster time to insight. In this paper, we discuss the lessons learned while supporting interactive, on-demand high performance computing from the perspectives of the users and the team supporting the users and the system. Building on these lessons, we present an overview of current needs and the technical solutions we are building to lower the barrier to entry for new users from the humanities, social, and biological sciences., Comment: 15 pages, 3 figures, First Workshop on Interactive High Performance Computing (WIHPC) 2018 held in conjunction with ISC High Performance 2018 in Frankfurt, Germany

Simulation, machine learning, and data analysis require a wide range of software which can be dependent upon specific operating systems, such as Microsoft Windows. Running this software interactively on massively parallel supercomputers can present many challenges. Traditional methods of scaling Microsoft Windows applications to run on thousands of processors have typically relied on heavyweight virtual machines that can be inefficient and slow to launch on modern manycore processors. This paper describes a unique approach using the Lincoln Laboratory LLMapReduce technology in combination with the Wine Windows compatibility layer to rapidly and simultaneously launch and run Microsoft Windows applications on thousands of cores on a supercomputer. Specifically, this work demonstrates launching 16,000 Microsoft Windows applications in 5 minutes running on 16,000 processor cores. This capability significantly broadens the range of applications that can be run at large scale on a supercomputer.

The Spectre and Meltdown flaws in modern microprocessors represent a new class of attacks that have been difficult to mitigate. The mitigations that have been proposed have known performance impacts. The reported magnitude of these impacts varies depending on the industry sector and expected workload characteristics. In this paper, we measure the performance impact on several workloads relevant to HPC systems. We show that the impact can be significant on both synthetic and realistic workloads. We also show that the performance penalties are difficult to avoid even in dedicated systems where security is a lesser concern.

Interactive massively parallel computations are critical for machine learning and data analysis. These computations are a staple of the MIT Lincoln Laboratory Supercomputing Center (LLSC) and has required the LLSC to develop unique interactive supercomputing capabilities. Scaling interactive machine learning frameworks, such as TensorFlow, and data analysis environments, such as MATLAB/Octave, to tens of thousands of cores presents many technical challenges - in particular, rapidly dispatching many tasks through a scheduler, such as Slurm, and starting many instances of applications with thousands of dependencies. Careful tuning of launches and prepositioning of applications overcome these challenges and allow the launching of thousands of tasks in seconds on a 40,000-core supercomputer. Specifically, this work demonstrates launching 32,000 TensorFlow processes in 4 seconds and launching 262,000 Octave processes in 40 seconds. These capabilities allow researchers to rapidly explore novel machine learning architecture and data analysis algorithms., Comment: 6 pages, 7 figures, IEEE High Performance Extreme Computing Conference 2018

The rise in computing hardware choices is driving a reevaluation of operating systems. The traditional role of an operating system controlling the execution of its own hardware is evolving toward a model whereby the controlling processor is distinct from the compute engines that are performing most of the computations. In this context, an operating system can be viewed as software that brokers and tracks the resources of the compute engines and is akin to a database management system. To explore the idea of using a database in an operating system role, this work defines key operating system functions in terms of rigorous mathematical semantics (associative array algebra) that are directly translatable into database operations. These operations possess a number of mathematical properties that are ideal for parallel operating systems by guaranteeing correctness over a wide range of parallel operations. The resulting operating system equations provide a mathematical specification for a Tabular Operating System Architecture (TabulaROSA) that can be implemented on any platform. Simulations of forking in TabularROSA are performed using an associative array implementation and compared to Linux on a 32,000+ core supercomputer. Using over 262,000 forkers managing over 68,000,000,000 processes, the simulations show that TabulaROSA has the potential to perform operating system functions on a massively parallel scale. The TabulaROSA simulations show 20x higher performance as compared to Linux while managing 2000x more processes in fully searchable tables., Comment: 8 pages, 6 figures, accepted at IEEE HPEC 2018

The rise of graph analytic systems has created a need for new ways to measure and compare the capabilities of graph processing systems. The MIT/Amazon/IEEE Graph Challenge has been developed to provide a well-defined community venue for stimulating research and highlighting innovations in graph analysis software, hardware, algorithms, and systems. GraphChallenge.org provides a wide range of pre-parsed graph data sets, graph generators, mathematically defined graph algorithms, example serial implementations in a variety of languages, and specific metrics for measuring performance. Graph Challenge 2017 received 22 submissions by 111 authors from 36 organizations. The submissions highlighted graph analytic innovations in hardware, software, algorithms, systems, and visualization. These submissions produced many comparable performance measurements that can be used for assessing the current state of the art of the field. There were numerous submissions that implemented the triangle counting challenge and resulted in over 350 distinct measurements. Analysis of these submissions show that their execution time is a strong function of the number of edges in the graph, $N_e$, and is typically proportional to $N_e^{4/3}$ for large values of $N_e$. Combining the model fits of the submissions presents a picture of the current state of the art of graph analysis, which is typically $10^8$ edges processed per second for graphs with $10^8$ edges. These results are $30$ times faster than serial implementations commonly used by many graph analysts and underscore the importance of making these performance benefits available to the broader community. Graph Challenge provides a clear picture of current graph analysis systems and underscores the need for new innovations to achieve high performance on very large graphs., Comment: 7 pages, 6 figures; submitted to IEEE HPEC Graph Challenge. arXiv admin note: text overlap with arXiv:1708.06866

Massive power-law graphs drive many fields: metagenomics, brain mapping, Internet-of-things, cybersecurity, and sparse machine learning. The development of novel algorithms and systems to process these data requires the design, generation, and validation of enormous graphs with exactly known properties. Such graphs accelerate the proper testing of new algorithms and systems and are a prerequisite for success on real applications. Many random graph generators currently exist that require realizing a graph in order to know its exact properties: number of vertices, number of edges, degree distribution, and number of triangles. Designing graphs using these random graph generators is a time-consuming trial-and-error process. This paper presents a novel approach that uses Kronecker products to allow the exact computation of graph properties prior to graph generation. In addition, when a real graph is desired, it can be generated quickly in memory on a parallel computer with no-interprocessor communication. To test this approach, graphs with $10^{12}$ edges are generated on a 40,000+ core supercomputer in 1 second and exactly agree with those predicted by the theory. In addition, to demonstrate the extensibility of this approach, decetta-scale graphs with up to $10^{30}$ edges are simulated in a few minutes on a laptop., Comment: 8 pages, 6 figures, IEEE IPDPS 2018 Graph Algorithm Building Blocks (GABB) workshop

An important objective for analyzing real-world graphs is to achieve scalable performance on large, streaming graphs. A challenging and relevant example is the graph partition problem. As a combinatorial problem, graph partition is NP-hard, but existing relaxation methods provide reasonable approximate solutions that can be scaled for large graphs. Competitive benchmarks and challenges have proven to be an effective means to advance state-of-the-art performance and foster community collaboration. This paper describes a graph partition challenge with a baseline partition algorithm of sub-quadratic complexity. The algorithm employs rigorous Bayesian inferential methods based on a statistical model that captures characteristics of the real-world graphs. This strong foundation enables the algorithm to address limitations of well-known graph partition approaches such as modularity maximization. This paper describes various aspects of the challenge including: (1) the data sets and streaming graph generator, (2) the baseline partition algorithm with pseudocode, (3) an argument for the correctness of parallelizing the Bayesian inference, (4) different parallel computation strategies such as node-based parallelism and matrix-based parallelism, (5) evaluation metrics for partition correctness and computational requirements, (6) preliminary timing of a Python-based demonstration code and the open source C++ code, and (7) considerations for partitioning the graph in streaming fashion. Data sets and source code for the algorithm as well as metrics, with detailed documentation are available at GraphChallenge.org., Comment: To be published in 2017 IEEE High Performance Extreme Computing Conference (HPEC)

The rise of graph analytic systems has created a need for ways to measure and compare the capabilities of these systems. Graph analytics present unique scalability difficulties. The machine learning, high performance computing, and visual analytics communities have wrestled with these difficulties for decades and developed methodologies for creating challenges to move these communities forward. The proposed Subgraph Isomorphism Graph Challenge draws upon prior challenges from machine learning, high performance computing, and visual analytics to create a graph challenge that is reflective of many real-world graph analytics processing systems. The Subgraph Isomorphism Graph Challenge is a holistic specification with multiple integrated kernels that can be run together or independently. Each kernel is well defined mathematically and can be implemented in any programming environment. Subgraph isomorphism is amenable to both vertex-centric implementations and array-based implementations (e.g., using the GraphBLAS.org standard). The computations are simple enough that performance predictions can be made based on simple computing hardware models. The surrounding kernels provide the context for each kernel that allows rigorous definition of both the input and the output for each kernel. Furthermore, since the proposed graph challenge is scalable in both problem size and hardware, it can be used to measure and quantitatively compare a wide range of present day and future systems. Serial implementations in C++, Python, Python with Pandas, Matlab, Octave, and Julia have been implemented and their single threaded performance have been measured. Specifications, data, and software are publicly available at GraphChallenge.org.

Increasing amounts of data from varied sources, particularly in the fields of machine learning and graph analytics, are causing storage requirements to grow rapidly. A variety of technologies exist for storing and sharing these data, ranging from parallel file systems used by supercomputers to distributed block storage systems found in clouds. Relatively few comparative measurements exist to inform decisions about which storage systems are best suited for particular tasks. This work provides these measurements for two of the most popular storage technologies: Lustre and Amazon S3. Lustre is an open-source, high performance, parallel file system used by many of the largest supercomputers in the world. Amazon's Simple Storage Service, or S3, is part of the Amazon Web Services offering, and offers a scalable, distributed option to store and retrieve data from anywhere on the Internet. Parallel processing is essential for achieving high performance on modern storage systems. The performance tests used span the gamut of parallel I/O scenarios, ranging from single-client, single-node Amazon S3 and Lustre performance to a large-scale, multi-client test designed to demonstrate the capabilities of a modern storage appliance under heavy load. These results show that, when parallel I/O is used correctly (i.e., many simultaneous read or write processes), full network bandwidth performance is achievable and ranged from 10 gigabits/s over a 10 GigE S3 connection to 0.35 terabits/s using Lustre on a 1200 port 10 GigE switch. These results demonstrate that S3 is well-suited to sharing vast quantities of data over the Internet, while Lustre is well-suited to processing large quantities of data locally., Comment: 5 pages, 4 figures, to appear in IEEE HPEC 2017

The MIT SuperCloud Portal Workspace enables the secure exposure of web services running on high performance computing (HPC) systems. The portal allows users to run any web application as an HPC job and access it from their workstation while providing authentication, encryption, and access control at the system level to prevent unintended access. This capability permits users to seamlessly utilize existing and emerging tools that present their user interface as a website on an HPC system creating a portal workspace. Performance measurements indicate that the MIT SuperCloud Portal Workspace incurs marginal overhead when compared to a direct connection of the same service., Comment: 6 pages, 3 figures, to appear in IEEE HPEC 2017