Dragonfly-based networks are an extensively deployed network topology in large-scale high-performance computing due to their cost-effectiveness and efficiency. The US will soon have three Exascale supercomputers for leadership class workloads deployed using dragonfly networks. Compared to indirect networks of similar scale, the dragonfly network has considerably reduced cable lengths, cable counts, and switch counts, resulting in significant network cost savings for a given system size, however, these cost reductions result in reduced global minimal paths and more challenging routing. Additionally, large scale dragonfly networks often require a taper at the global link level, resulting in less bisection bandwidth than is achievable in other traditional non-blocking topologies of equivalent scale. While dragonfly networks have been extensively studied, they have yet to be fully evaluated in an extreme scale (i.e., exascale) system that targets capability workloads. In this paper, we present the results of the first large scale evaluation of a dragonfly network on an exascale system (Frontier) and compare its behavior to a similar scale fat-tree network on a previous generation TOP500 system (Summit). This evaluation aims to determine the effect of network cost optimizations by measuring a tapered topology’s impact on capability workloads. Our evaluation is based on a collection of synthetic microbenchmarks, mini-apps, and full scale applications. It compares the scaling efficiencies of each benchmark between the dragonfly-based Frontier and the fat-tree-based Summit systems. Our results show that a dragonfly network is $\sim \mathbf{3 0 \%}$ more cost efficient than a fat-tree topology, which amortizes to $\sim 3 \%$ of an exascale system cost. Furthermore, while tapered dragonfly networks impose significant tradeoffs, the impacts are not as broad as initially thought and are mostly seen in applications with global communication patterns, particularly all-to-all (e.g., FFT-based algorithms), but also local communication patterns (e.g., nearest-neighbor algorithms) that are sensitive to network performance variability.

As the US Department of Energy (DOE) computing facilities began deploying petascale systems in 2008, DOE was already setting its sights on exascale. In that year, DARPA published a report on the feasibility of reaching exascale. The report authors identified several key challenges in the pursuit of exascale including power, memory, concurrency, and resiliency. That report informed the DOE's computing strategy for reaching exascale. With the deployment of Oak Ridge National Laboratory's Frontier supercomputer, we have officially entered the exascale era. In this paper, we discuss Frontier's architecture, how it addresses those challenges, and describe some early application results from Oak Ridge Leadership Computing Facility's Center of Excellence and the Exascale Computing Project.

Scientific communities are increasingly adopting deep learning (DL) models in their applications to accelerate scientific discovery processes. However, with rapid growth in the computing capabilities of HPC supercomputers, large-scale DL applications have to spend a significant portion of training time performing I/O to a parallel storage system. Previous research works have investigated optimization techniques such as prefetching and caching. Unfortunately, there exist non-trivial challenges to adopting the existing solutions on HPC supercomputers for large-scale DL training applications, which include non-performance and/or failures at extreme scale, lack of portability and generality in design, complex deployment methodology, and being limited to a specific application or dataset. To address these challenges, we propose High-Velocity AI Cache (HVAC), a distributed read-cache layer that targets and fully exploits the node-local storage or near node-local storage technology. HVAC seamlessly accelerates read I/O by aggregating node-local or near node-local storage, avoiding metadata lookups and file locking while preserving portability in the application code. We deploy and evaluate HVAC on 1,024 nodes (with over 6000 NVIDIA V100 GPUS) of the Summit supercomputer. In particular, we evaluate the scalability, efficiency, accuracy, and load distribution of HVAC compared to GPFS and XFS-on-NVMe. With four different DL applications, we observe an average 25 % performance improvement atop GPFS and 9% drop against XFS-on-NVMe, which scale linearly and are considered the performance upper bound. We envision HVAC as an important caching library for upcoming HPC supercomputers such as Frontier.

Burst buffers (BBs) are increasingly exploited in contemporary supercomputers to bridge the performance gap between compute and storage systems. The design of BBs, particularly the placement of these devices and the underlying network topology, impacts both performance and cost. As the cost of other components such as memory and accelerators is increasing, it is becoming more important that HPC centers provision BBs tailored to their workloads.This work contributes a provisioning system to provide accurate, multi-tenant simulations that model realistic application and storage workloads from HPC systems. The framework aids HPC centers in modeling their workloads against multiple network and BB configurations rapidly. In experiments with our framework, we provide a comparison of representative Oak Ridge Leadership Computing Facility (OLCF) I/O workloads against multiple BB designs. We analyze the impact of these designs on latency, I/O phase lengths, contention for network and storage devices, and choice of network topology.

The I/O subsystem for the Summit supercomputer, No. 1 on the Top500 list, and its ecosystem of analysis platforms is composed of two distinct layers, namely the in-system layer and the center-wide parallel file system layer (PFS), Spider 3. The in-system layer uses node-local SSDs and provides 26.7 TB/s for reads, 9.7 TB/s for writes, and 4.6 billion IOPS to Summit. The Spider 3 PFS layer uses IBM's Spectrum Scale™ and provides 2.5 TB/s and 2.6 million IOPS to Summit and other systems. While deploying them as two distinct layers was operationally efficient, it also presented usability challenges in terms of multiple mount points and lack of transparency in data movement. To address these challenges, we have developed novel end-to-end I/O solutions for the concerted use of the two storage layers. We present the I/O subsystem architecture, the end-to-end I/O solution space, their design considerations and our deployment experience.

The US Department of Energy deployed the Summit and Sierra supercomputers with the latest state-of-the-art network interconnect technology in 2018 and both systems entered production in 2019. In this paper, we provide an in-depth assessment of the systems' network interconnects that are based on Enhanced Data Rate (EDR) 100 Gb/s Mellanox InfiniBand. Both systems use second-generation EDR Host Channel Adapters (HCAs) and switches with several new features such as Adaptive Routing (AR), switch-based collectives, and HCA-based tag matching. Although based on the same components, Summit's network is "non-blocking" (i.e., a fully provisioned Clos network) and Sierra's network has a 2:1 taper between the racks and aggregation switches. We evaluate the two systems' interconnects using traditional communication benchmarks as well as production applications. We find that the new Adaptive Routing dramatically improves performance but the other new features still need improvement.

While future terabit networks hold the promise of significantly improving big-data motion among geographically distributed data centers, significant challenges must be overcome even on today's 100 gigabit networks to realize end-to-end performance. Multiple bottlenecks exist along the end-to-end path from source to sink, for instance, the data storage infrastructure at both the source and sink and its interplay with the wide-area network are increasingly the bottleneck to achieving high performance. In this paper, we identify the issues that lead to congestion on the path of an end-to-end data transfer in the terabit network environment, and we present a new bulk data movement framework for terabit networks, called LADS. LADS exploits the underlying storage layout at each endpoint to maximize throughput without negatively impacting the performance of shared storage resources for other users. LADS also uses the Common Communication Interface (CCI) in lieu of the sockets interface to benefit from hardware-level zero-copy, and operating system bypass capabilities when available. It can further improve data transfer performance under congestion on the end systems using buffering at the source using flash storage. With our evaluations, we show that LADS can avoid congested storage elements within the shared storage resource, improving input/output bandwidth, and data transfer rates across the high speed networks. We also investigate the performance degradation problems of LADS due to I/O contention on the parallel file system (PFS), when multiple LADS tools share the PFS. We design and evaluate a meta-scheduler to coordinate multiple I/O streams while sharing the PFS, to minimize the I/O contention on the PFS. With our evaluations, we observe that LADS with meta-scheduling can further improve the performance by up to 14 percent relative to LADS without meta-scheduling.

Increased growth of the data and the need to move the data between data centers, demands, an efficient data transfer tool which can, not only transfer the data at higher rates but also handle the faults occurred during the transfer. Absence of fault tolerance mechanisms, would need to retransmit the whole data, in case of any error during the transfer. Hence, fault tolerance is an important aspect of big data transfer tools. In this paper, we have considered LADS data transfer tool, which proved to be superior to the existing data transfer tools with respect to the speed of the transfer. However, absence of fault tolerance implementation in LADS might result in data retransmission and congestion issues upon errors. In this paper, we design and analyze fault tolerance mechanisms which can be used with LADS data transfer tool. We have proposed three different fault tolerance mechanisms, File logging, Transaction logging and Universal logging. Also, we have analyzed the space and performance overhead of these fault tolerance mechanisms on LADS data transfer tool.

Job placement plays a pivotal role in application performance on supercomputers. We present a multi-faceted exploration to influence placement in extreme-scale systems, to improve network performance and decrease variability. In our first exploration, Scores, we developed a machine learning model that extracts features from a job's node-allocation and grades performance. This identified several important node-metrics that led to Dual-Ended scheduling, a means of reducing network contention without impacting utilization. In evaluations on the Titan supercomputer, we observed reductions in average hop-count by up to 50%. We also developed an improved node-layout strategy that targets a better balance between network latency and bandwidth, replacing the default ALPS layout on Titan that resulted in an average of 10% runtime improvement. Both of these efforts underscore the importance of a job placement strategy that is cognizant of workload mixture and network topology.

The growth of computing power on large-scale systems requires commensurate high-bandwidth I/O systems. Many parallel file systems are designed to provide fast sustainable I/O in response to applications' soaring requirements. To meet this need, a novel system is imperative to temporarily buffer the bursty I/O and gradually flush datasets to long-term parallel file systems. In this paper, we introduce the design of BurstMem, a high-performance burst buffer system. BurstMem provides a storage framework with efficient storage and communication management strategies. Our experiments demonstrate that BurstMem is able to speed up the I/O performance of scientific applications by up to 8.5× on leadership computer systems.

Many science facilities, such as the Department of Energy's Leadership Computing Facilities and experimental facilities including the Spallation Neutron Source, Stanford Linear Accelerator Center, and Advanced Photon Source, produce massive amounts of experimental and simulation data. These data are often shared among the facilities and with collaborating institutions. Moving large datasets over the wide-area network (WAN) is a major problem inhibiting collaboration. Next-generation, terabit-networks will help alleviate the problem, however, the parallel storage systems on the endsystem hosts at these institutions can become a bottleneck for terabit data movement. The parallel storage system (PFS) is shared by simulation systems, experimental systems, analysis and visualization clusters, in addition to wide-area data movers. These competing uses often induce temporary, but significant, I/O load imbalances on the storage system, which impact the performance of all the users. The problem is a serious concern because some resources are more expensive (e.g. super computers) or have time-critical deadlines (e.g. experimental data from a light source), but parallel file systems handle all requests fairly even if some storage servers are under heavy load. This paper investigates the problem of competing workloads accessing the parallel file system and how the performance of wide-area data movement can be improved in these environments. First, we study the I/O load imbalance problems using actual I/O performance data collected from the Spider storage system at the Oak Ridge Leadership Computing Facility. Second, we present I/O optimization solutions with layout-awareness on end-system hosts for bulk data movement. With our evaluation, we show that our I/O optimization techniques can avoid the I/O congested disk groups, improving storage I/O times on parallel storage systems for terabit data movement.

Scientific applications often need to access remotely located files, but many remote I/O systems lack standard APIs that allow efficient and direct access from application codes. This work presents MPI-IO/L, a remote I/O facility for MPI-IO using logistical networking. This combination not only provides high-performance and direct remote I/O using the standard parallel I/O interface but also offers convenient management and sharing of remote files. We show the performance trade-offs with various remote I/O approaches implemented in the system, which can help scientists identify preferable I/O options for their own applications. We also discuss how logistical networking could be improved to work better with parallel I/O systems such as ROMIO.

We propose a distributed data management scheme for large data visualization that emphasizes efficient data sharing and access. To minimize data access time and support users with a variety of local computing capabilities, we introduce an adaptive data selection method based on an "enhanced time-space partitioning" (ETSP) tree that assists with effective visibility culling, as well as multiresolution data selection. By traversing the tree, our data management algorithm can quickly identify the visible regions of data, and, for each region, adaptively choose the lowest resolution satisfying user-specified error tolerances. Only necessary data elements are accessed and sent to the visualization pipeline. To further address the issue of sharing large-scale data among geographically distributed collaborative teams, we have designed an infrastructure for integrating our data management technique with a distributed data storage system provided by logistical networking (LoN). Data sets at different resolutions are generated and uploaded to LoN for wide-area access. We describe a parallel volume rendering system that verifies the effectiveness of our data storage, selection and access scheme.