Engineering topics

Dorier, Matthieu

Publications and source records attributed to Dorier, Matthieu.

At least 19 records

DOE OSTI2024

Diaspora: Resilience-Enabling Services for Real-Time Distributed Workflows

The need for real-time processing to enable automated decision making and experimental steering has driven a shift from high-performance computing workflows on a centralized system to a distributed approach that integrates remote data sources, edge devices, and diverse compute facilities. Under this paradigm, data can be processed close to the source where it is generated, thus reducing latency and bandwidth usage. System resilience is thus a key challenge, requiring distributed workflows to survive component failures and to meet stringent quality-of-service requirements, which results in the need to mitigate anomalies such as congestion and low availability of resources. To address these challenges, we propose Diaspora, a unified resilience framework that is inspired by event-driven communication patterns used in public clouds. Specifically, we propose an event fabric that extends across sites, facilities, and computations to provide timely, reliable, and accurate information about data, application, and resource status. On top of the event fabric, we build resilience-enabling services that combine QoS-aware data streaming, resilient data views, resilient compute and data resources, and anomaly detection and prediction, all of which collectively enhance workflow resilience for these scientific cases.

Rao, Nageswara↗

DOE OSTI2024

Performance Characterization and Provenance of Distributed Task-based Workflows on HPC Platforms

Understanding performance and provenance of task-based workflows poses significant challenges, particularly in distributed configurations where resources are shared by multiple applications. Task-based workflow management systems further complicate performance predictability because of their dynamicity that subtly alters task execution order from run to run. In this paper we propose a layered characterization framework for performance and task provenance for Dask.distributed workflows running on high-performance computing (HPC) platforms. It collects data from jobs, the workflow management system, and the operating system to aid in understanding the performance of these workflows. Our approach encompasses three main contributions: first, an extension of Dask.distributed to capture high-fidelity task provenance using Mochi data services; second, the adaptation of the established HPC I/O characterization tool Darshan to gather high-fidelity I/O data, thereby enhancing the granularity of our analysis; and third, a framework to combine and process the collected data and provide helpful insights into performance characterization and reproducibility, alongside our lessons learned.

Dask↗

DOE OSTI2023

HEPnOS: a Specialized Data Service for High Energy Physics Analysis

In this paper, we present HEPnOS, a distributed data service for managing data produced by high-energy physics (HEP) experiments. Using HEPnOS, HEP applications can use HPC resources more effciently than traditional fle-based applications. The fle-based model leads to a rigid, chunk-based allocation of computational resources and limits the number of cores that can be used concurrently by an HEP application. The fundamental problem is that organizing domain-specifc data into fles inadvertently introduces a single, artifcial, confated tuning parameter that puts key optimization goals into confict: larger fle sizes reduce metadata overhead and thus improve I/O effciency, but smaller fle sizes provide more opportunity for workfow parallelism and load balancing. In this work, we introduce a domain-specifc data service that decouples that constraint so that data can be accessed and processed in its natural granularity while still maintaining I/O effciency. By removing the constraints introduced by fle handling we are able to obtain better scaling and make effcient use of more cores for processing a fxed-sized data sample. We demonstrate the improved scalability by using an application developed in the fle-based paradigm and comparing it to a version modifed to use HEPnOS.

72 PHYSICS OF ELEMENTARY PARTICLES AND FIELDS↗

DOE OSTI2023

Towards elastic in situ analysis for high-performance computing simulations

In situ analysis and visualization have grown increasingly popular for enabling direct access to data from high-performance computing (HPC) simulations. As a simulation progresses and interesting physical phenomena emerge, however, the data produced may become increasingly complex, and users may need to dynamically change the type and scale of in situ analysis tasks being carried out and consequently adapt the amount of resources allocated to such tasks. To date, none of the production in situ analysis frameworks offer such an elasticity feature, and for good reason: the assumption that the number of processes could vary during run time would force developers to rethink software and algorithms at every level of the in situ analysis stack. In this paper we present Colza, a data staging service with elastic in situ visualization capabilities. We demonstrate the use of Colza with the Deep Water Impact and the AMR-Wind simulations, coupling them with the ParaView Catalyst and Ascent in situ libraries, and show that Colza enables dynamic rescaling of these widely-used frameworks with no interruption to the simulation or staging service. Here, we highlight the challenges of enabling such elasticity, which requires overcoming these frameworks' reliance on MPI, using distinct engineering approaches, namely dependency injection and dependency overload. To the best of our knowledge, this work is the first to enable elastic in situ visualization capabilities for HPC applications on top of existing production analysis tools.

97 MATHEMATICS AND COMPUTING↗

DOE OSTI2022

Adaptive elasticity policies for staging-based in situ visualization

In situ processing aims to alleviate the growing gap between computation and I/O capabilities by performing data processing close to the data source. In situ processing is widely used to process data generated by multiple data sources, including observation data from edge devices or scientific observational facilities and the simulation data generated by scientific computation on a high-performance computing (HPC) platform. For a scientific workflow that is run on an HPC platform and composed of a simulation program and an in situ data analytics or visualization (abbreviated as ana/vis) task, there is an implicit assumption that the computing resources assigned to the workflow keep static during the workflow execution. However, with the converging trend between the HPC and cloud computing platform, running the in situ ana/vis task in an elastic way is promising to decrease its overhead and improve its resource utilization rate. Resource elasticity represents the ability to change resource configurations such as the number of computing nodes/processes during workflow execution. An elastic job may dynamically adjust resource configurations; it may use a few resources at the beginning and more resources toward the end of the job when interesting data appear. However, it is hard to predict a priori how many computing nodes/processes need to be added/removed during the workflow execution to adapt to changing workflow needs. How to efficiently guide elasticity operations, such as growing or shrinking the number of processes used for in situ analysis during workflow execution, is an open-ended research question. In this article, we present adaptive elasticity policies that adopt workflow runtime information collected during workflow execution to predict how to trigger the addition/removal of processes in order to minimize in situ processing overhead. Taking in situ visualization tasks as an example, we integrate the presented elasticity policies into a staging-based elastic workflow and evaluate its efficiency in multiple elasticity scenarios. Compared with the situation without elasticity or with a static elasticity policy that uses a fixed number of processes for each rescaling operation, the adaptive elasticity policy can save overhead in finding a proper resource configuration and improve resource utilization efficiency. Furthermore, one experiment illustrates that the adaptive elasticity policy saves 41% of core-hours compared with the situation without the resource elasticity.

97 MATHEMATICS AND COMPUTING↗

DOE OSTI2021

SYMBIOSYS: A Methodology for Performance Analysis of Composable HPC Data Services

Microservices are a powerful new way of building, customizing, and deploying distributed services owing to their flexibility and maintainability. Several large-scale distributed platforms have emerged to serve the growing needs of data-centric workloads and services in commercial computing. Concurrently, high-performance computing (HPC) systems and software are rapidly evolving to meet the demands of diversified applications and heterogeneity. The interplay of hardware factors, software configuration parameters, and the flexibility offered with a microservice architecture makes it nontrivial to estimate the optimal service instantiation for a given application workload. Further, this problem is exacerbated when considering that these services operate in a dynamic and heterogeneous HPC environment. An optimally integrated service can be vastly more performant than a haphazardly integrated one. Existing performance tools for HPC either fail to understand the request-response model of communication inherent to microservices or they operate within a narrow scope, limiting the insight that can be gleaned from employing them in isolation. We propose a methodology for integrated performance analysis of HPC microservices frameworks and applications called SYMBIOSYS. We describe its design and implementation within the context of the Mochi framework. This integration is achieved by combining distributed callpath profiling and tracing with a performance data exchange strategy that collects fine-grained, low-level metrics from the RPC communication library and network layers. The result is a portable, low-overhead performance analysis setup that provides a holistic profile of the dependencies among microservices and how they interact with the Mochi RPC software stack. Using HEPnOS, a production-quality Mochi data service, we demonstrate the low-overhead operation of SYMBIOSYS at scale and use it to identify the root causes of poorly performing service configurations.

microservices↗

DOE OSTI2021

Facilitating Staging-based Unstructured Mesh Processing to Support Hybrid In-Situ Workflows

In-situ and in-transit processing alleviate the gap between the computing and I/O capabilities by scheduling data analytics close to the data source. Hybrid in-situ processing splits data analytics into two stages: the data processing that runs in-situ aims to extract regions of interest, which are then transferred to staging services for further in-transit analytics. To facilitate this type of hybrid in-situ processing, the data staging service needs to support complex intermediate data representations generated/consumed by the in-situ tasks. Unstructured (or irregular) mesh is one such derived data representation that is typically used and bridges simulation data and analytics. However, how staging services efficiently support unstructured mesh transfer and processing remains to be explored. This paper investigates design options for transferring and processing unstructured mesh data using staging services. Using polygonal mesh data as an example, we show that hybrid in-situ workflows with staging-based unstructured mesh processing can effectively support hybrid in-situ workflows, and can significantly decrease data movement overheads.

data-driven↗

DOE OSTI2020

A terminology for in situ visualization and analysis systems

The term "in situ processing" has evolved over the last decade to mean both a specific strategy for visualizing and analyzing data and an umbrella term for a processing paradigm. The resulting confusion makes it difficult for visualization and analysis scientists to communicate with each other and with their stakeholders. To address this problem, a group of over 50 experts convened with the goal of standardizing terminology. This paper summarizes their findings and proposes a new terminology for describing in situ systems. An important finding from this group was that in situ systems are best described via multiple, distinct axes: integration type, proximity, access, division of execution, operation controls, and output type. This paper discusses these axes, evaluates existing systems within the axes, and explores how currently used terms relate to the axes.

97 MATHEMATICS AND COMPUTING↗

DOE OSTI2020

How fast can one resize a distributed file system?

Efficient resource utilization becomes a major concern as large-scale distributed computing infrastructures keep growing in size. Malleability, the possibility for resource managers to dynamically increase or decrease the amount of resources allocated to a job, is a promising way to save energy and costs. However, state-of-the-art parallel and distributed storage systems have not been designed with malleability in mind. The reason is mainly the supposedly high cost of data transfers required by resizing operations. Nevertheless, as network and storage technologies evolve, old assumptions about potential bottlenecks can be revisited. In this study, we evaluate the viability of malleability as a design principle for a distributed storage system. We specifically model the minimal duration of the commission and decommission operations. To show how our models can be used in practice, we evaluate the performance of these operations in HDFS, a relevant state-of-the-art distributed file system. We show that the existing decommission mechanism of HDFS is good when the network is the bottleneck, but can be accelerated by up to a factor 3 when storage is the limiting factor. We also show that the commission in HDFS can be substantially accelerated. With the highlights provided by our model, we suggest improvements to speed both operations in HDFS. We discuss how the proposed models can be generalized for distributed file systems with different assumptions and what perspectives are open for the design of efficient malleable distributed file systems.

97 MATHEMATICS AND COMPUTING↗

DOE OSTI2020

Mochi: Composing Data Services for High-Performance Computing Environments

Technology enhancements and the growing breadth of application workflows running on high-performance computing (HPC) platforms drive the development of new data services that provide high performance on these new platforms, provide capable and productive interfaces and abstractions for a variety of applications, and are readily adapted when new technologies are deployed. The Mochi framework enables composition of specialized distributed data services from a collection of connectable modules and subservices. Rather than forcing all applications to use a one-size-fits-all data staging and I/O software configuration, Mochi allows each application to use a data service specialized to its needs and access patterns. This paper introduces the Mochi framework and methodology. The Mochi core components and microservices are described. Examples of the application of the Mochi methodology to the development of four specialized services are detailed. Finally, a performance evaluation of a Mochi core component, a Mochi microservice, and a composed service providing an object model is performed. The paper concludes by positioning Mochi relative to related work in the HPC space and indicating directions for future work.

97 MATHEMATICS AND COMPUTING↗

DOE OSTI2020

Pufferscale: Rescaling HPC Data Services for High Energy Physics Applications

User-space HPC data services are emerging as an appealing alternative to traditional parallel file systems, because of their ability to be tailored to application needs while eliminating unnecessary overheads incurred by POSIX compliance. The High Energy Physics (HEP) community is progressively turning towards such services to enable high-throughput accesses under heavy concurrency to billions of event data produced by instruments and consumed by subsequent analysis workflows. Such services would benefit from the possibility to be rescaled up and down to adapt to changing workloads, as experimental campaigns progress, in order to optimize resource usage. This paper formalizes rescaling a distributed storage system as a multi objective optimization problem considering three criteria: load balance, data balance, and duration of the rescaling operation. We propose a heuristic for rapidly finding a good approximate solution, while allowing users to weight the criteria as needed. The heuristic is evaluated with Pufferscale, a new rescaling manager for microservice-based distributed storage systems. To validate our approach in a real-world ecosystem, we showcase the use of Pufferscale as a means to enable storage malleability in the HEPnOS storage system for HEP applications.

Distributed Storage System↗

DOE OSTI2020

DeepFreeze: Towards Scalable Asynchronous Checkpointing of Deep Learning Models

In the age of big data, deep learning has emerged as a powerful tool to extract insight and exploit its value, both in industry and scientific applications. One common pattern emerging in such applications is frequent checkpointing of the state of the learning model during training, needed in a variety of scenarios: analysis of intermediate states to explain features and correlations with training data, exploration strategies involving alternative models that share a common ancestor, knowledge transfer, resilience, etc. However, with increasing size of the learning models and popularity of distributed data-parallel training approaches, simple checkpointing techniques used so far face several limitations: low serialization performance, blocking I/O, stragglers due to the fact that only a single process is involved in checkpointing. This paper proposes a checkpointing technique specifically designed to address the aforementioned limitations, introducing efficient asynchronous techniques to hide the overhead of serialization and I/O, and distribute the load over all participating processes. Experiments with two deep learning applications (CANDLE and ResNet) on a pre-Exascale HPC platform (Theta) shows significant improvement over state-of-art, both in terms of checkpointing duration and runtime overhead.

checkpointing↗

DOE OSTI2020

DeepClone: Lightweight State Replication of Deep Learning Models for Data Parallel Training

Training modern deep neural network (DNN) models involves complex workflows triggered by model exploration, sensitivity analysis, explainability, etc. A key primitive in this context is the ability to clone a model training instance, i.e. “fork” the training process in a potentially different direction, which enables comparisons of different evolution paths using variations of training data and model parameters. However, in a quest improve the training throughput, a mix of data parallel, model parallel, pipeline parallel and layer-wise parallel approaches are making the problem of cloning highly complex. In this paper, we explore the problem of efficient cloning under such circumstances. To this end, we leverage several properties of data-parallel training and layer-wise parallelism to design DeepClone, a cloning approach based on augmenting the execution graph to gain direct access to tensors, which are then sharded and reconstructed asynchronously in order to minimize runtime overhead, standby duration, readiness duration. Compared with state-of-art approaches, DeepClone shows orders of magnitude improvement for several classes of DNN models.

data-parallel training↗

DOE OSTI2019

A Software Defined Storage Approach to Exascale Storage Services (Final Technical Report)

The Mochi project began in 2015 in recognition of a set of changes that were set to alter the landscape of storage and I/O for DOE science. On the one hand, an influx of new classes of applications including an initial set of data intensive and machine learning applications were operating with new types of data and in new ways, as compared to traditional computational science simulation codes. Simultaneously, new memory, storage, and networking technologies were driving a need to adapt software designs to a rapidly changing design space. The goal of the three-year Mochi project was to develop and demonstrate an approach for composing storage services which provides new levels of functionality, performance, and reliability for science applications at extreme scale. The team, including a mix of Laboratory, university, and industry collaborators, focused on systems and applications of interest to DOE Office of Science, tracking developments at NERSC and the LCFs as well as application trends within ECP and SciDAC. Technical work focused on three themes: composition (understand the building blocks needed to provide storage services that meet DOE mission needs and how these components are flexibly composed), application interfacing (what service compositions are most effective for common DOE application use cases), and consistency and synchronization (taking advantage of knowledge about, and from, the system and scientific applications to provide better service than would otherwise be possible). The project was highly successful, producing 13 publications and tens of thousands of lines of open source software that is now being used by multiple ECP teams, HPC software vendors, and international collaborators. The uptake of Mochi technologies speaks to both the value of the service composition model and the quality of architecture and software developed by the team. Through continued funding under ASCR and ECP, the Mochi team and its technologies are on track to further contribute to DOE science successes into the exascale timeframe.

97 MATHEMATICS AND COMPUTING↗

DOE OSTI2019

MPI jobs within MPI jobs: A practical way of enabling task-level fault-tolerance in HPC workflows

While the use of workflows for HPC is growing, MPI interoperability remains a challenge for workflow management systems. The MPI standard and/or its implementations provide a number of ways to build multiple-programs-multiple-data (MPMD) applications. These methods present limitations related to fault tolerance, and are not easy to use. In this paper, we advocate for a novel MPI_Comm_launch function acting as the parallel counterpart of a system(3) call. MPI_Comm_launch allows a child MPI application to be launched inside the resources originally held by processes of a parent MPI application. Two important aspects of MPI_Comm_launch is that it pauses the calling process, and runs the child processes on the parent’s CPU cores, but in an isolated manner with respect to memory. This function makes it easier to build MPMD applications with well-decoupled subtasks. We show how this feature can provide better flexibility and better fault tolerance in ensemble simulations and HPC workflows. Here, we report results showing 2x throughput improvement for application workflows with faults, and scaling results for challenging workloads up to 256 nodes.

97 MATHEMATICS AND COMPUTING↗

DOE OSTI2019

Is it Worth Relaxing Fault Tolerance to Speed Up Decommission in Distributed Storage Systems?

Efficient resource utilization is a major concern for large-scale computer platforms. One method used to lower energy consumption and operational cost is to reduce the amount of idle resources. This can be achieved by using malleability, namely, the possibility for resource managers to dynamically increase or decrease the amount of resources of jobs while they are running. Decommissioning (i.e., removing from the cluster) the idle nodes as soon as possible allows the resource manager to quickly reallocate those nodes to other jobs. Challenges appear when such nodes host part of a distributed storage system. Such storage systems may need to transfer large amounts of data before releasing the nodes, in order to ensure data availability and a certain level of fault tolerance. In this paper, we model and evaluate the performance of the decommission operation when relaxing the level of fault tolerance (i.e., the number of replicas) during this operation. Intuitively, this is expected to reduce the amount of data transfers needed before nodes are released, and thus allow nodes to be returned to the resource manager faster. We quantify theoretically how much time and resources are saved by such a fast decommission strategy compared with a standard decommission that does not temporarily reduce the fault-tolerance level. We establish lower bounds for the duration of the different phases of a fast decommission. We use the lower bounds to estimate when fast decommission would be useful to reduce the usage of core-hours and when not. We implement a prototype for fast decommission and experimentally validate the lower bounds on the duration of the operation and confirm in practice our theoretical findings.

Decommission↗

DOE OSTI2018

Supporting HEP Data as the Exascale Era Approaches

We describe our work on the use of HPC data formats and storage technologies to efficiently describe, store and access High Energy Physics (HEP) data. In the HEP community, important data sets are large, currently at the scale of tens of PiB, and are expected to grow dramatically in the next decade. These data require different data representations—ranging from complex C++ objects to tabular data—that are used at the many varied stages of data analysis. Traditionally, these data have been stored in large collections of files at each stage. This approach does not scale well on HPC resources. To obtain good scalability, without requiring that analysts have parallel programming expertise, our work stresses the use of HPCoptimized tools and implicit data-parallelism. We describe data organization and the performance and scaling of data processing using examples from three HEP experiments (LArIAT, NOvA and DUNE), and for three different stages of analysis. We compare the performance of our HPC-oriented solutions to the traditional HEP approaches.

Buchanan, Norman↗

DOE OSTI2017

Yokan

Remote Key/Value storage service for Mochi.

Dorier, Matthieu↗