Modern GPU-rich HPC systems are increasingly becoming energy-constrained. Thus, understanding an application’s energy consumption becomes essential. Unfortunately, current GPU energy attribution techniques are either inaccurate, inflexible, or outdated. Therefore, we propose Wattchmen, a flexible methodology for measuring, attributing, and predicting GPU energy consumption. We construct a per-instruction energy model using a diverse set of microbenchmarks to systematically quantify the energy consumption of GPU instructions, enabling finer-grain prediction and energy consumption breakdowns for applications. Compared with the state-of-the-art systems like AccelWattch (32%) and Guser (25%), across 16 popular GPGPU, graph analytics, HPC, and ML workloads, Wattchmen reduces the mean absolute percent error (MAPE) to 14% on V100 GPUs. Furthermore, we show that Wattchmen provides similar MAPEs for water-cooled V100s (15%) and extends to later architectures, including air-cooled A100 (11%) and H100 (12%) GPUs. Finally, to further demonstrate Wattchmen ’s value, we apply it to applications such as Backprop and QMCPACK, where Wattchmen ’s insights enable energy reductions of up to 35%.

Electrical provisioning in high performance computing is transitioning from simple nameplate Thermal Design Power (TDP) models to more nuanced approaches based on expected electrical load.This paper captures current power

The power & energy demands of HPC machines have grown significantly. Modern exascale HPC systems require tens of megawatts of combined power for computing resources and cooling facilities at full capacity. The current energy trend is not sustainable for future HPC systems, and there is a need to work toward the energy efficiency aspect of HPC performance. Energy awareness of the HPC applications at the job level is essential for running an efficient HPC system. This work aims to develop a pipeline to provide a production-level system-wide overview of the HPC workloads' power profile while handling evolving workloads exhibiting new power trends. We developed an open-set classification model for HPC jobs based on the properties of power profiles to continuously provide a system-wide holistic view of recently completed jobs. The pipeline helps continuously monitor the job-level power usage pattern of HPC and enables us to capture the new trends in applications' power behavior. We employed a comprehensive set of techniques to generate job-level data, custom-designed feature extraction methods to extract critical features from jobs' power profiles, clustering techniques powered by generative modeling, and open-set classification for identifying job profiles into known classes or an unknown set. With extensive evaluations, we demonstrate the effectiveness of each component in our pipeline. We provide an analysis of the resulting clusters that characterize the power profile landscape of the Summit supercomputer from more than 60K jobs executed in a year. The open-set classification classifies the known data sets into known classes with high accuracy and identifies unknown data noints with over 85% accuracy.

GPU memory corruption and in particular double-bit errors (DBEs) remain one of the least understood aspects of HPC system reliability. Albeit rare, their occurrences always lead to job termination and can potentially cost thousands of node-hours, either from wasted computations or as the overhead from regular checkpointing needed to minimize the losses. As supercomputers and their components simultaneously grow in scale, density, failure rates, and environmental footprint, the efficiency of HPC operations becomes both an imperative and a challenge.

We present ExaDigiT, an open-source framework for developing comprehensive digital twins of liquid-cooled supercomputers. It integrates three main modules: (1) a resource allocator and power simulator, (2) a transient thermo-fluidic cooling model, and (3) an augmented reality model of the supercomputer and central energy plant. The framework enables the study of “what-if” scenarios, system optimizations, and virtual prototyping of future systems. Using Frontier as a case study, we demonstrate the framework’s capabilities by replaying six months of system telemetry for systematic verification and validation. Such a comprehensive analysis of a liquid-cooled exascale supercomputer is the first of its kind. ExaDigiT elucidates complex transient cooling system dynamics, runs synthetic or real workloads, and predicts energy losses due to rectification and voltage conversion. Throughout our paper, we present lessons learned to benefit HPC practitioners developing similar digital twins. We envision the digital twin will be a key enabler for sustainable, energy-efficient supercomputing.

In this work, we demonstrate the challenges in predicting HPC cluster power consumption in the face of significant temporal skew in power consumption behavioral patterns. Predicting large power swings that extend several megawatts has significant operational value for HPC centers, however, prediction is challenging due to the relative rarity of such events and also due to the abrupt or disjoint deviation from the average power consumption levels. To study the impact of this challenge, we have trained a recurrent neural network (RNN) as a reasonably sophisticated model to predict power consumption of the one-year worth of node power consumption data from the Summit supercomputer located in the Oak Ridge Leadership Computing Facility. By studying the prediction results, we have found that although simple usage of RNN models can provide good results on average power consumption levels, it would fail at predicting the power swings that have more operational value. With such results, we discuss potential next steps in addressing such issues aiming towards a robust usage of power prediction techniques in HPC operations.

This paper provides a broad framework for understanding trends in Operational Data Analytics (ODA) for High-Performance Computing (HPC) facilities. The goal of ODA is to allow for the continuous monitoring, archiving, and analysis of near real-time performance data, providing immediately actionable information for multiple operational uses. In this work, we combine two models to provide a comprehensive HPC ODA framework: one is an evolutionary model of analytics capabilities that consists of four types, which are descriptive, diagnostic, predictive and prescriptive, while the other is a four-pillar model for energy-efficient HPC operations that covers facility, system hardware, system software, and applications. This new framework is then overlaid with a description of current development and production deployments of ODA within leading-edge HPC facilities. Finally, we perform a comprehensive survey of ODA works and classify them according to our framework, in order to demonstrate its effectiveness.

As we approach the exascale computing era, the focused understanding of power consumption and its overall constraint on HPC architectures and applications are becoming increasingly paramount. Summit, located at the Oak Ridge Leadership Computing Facility (OLCF), is one of the fastest and largest pre-exascale platforms in operation today. This paper provides a first-order examination and analysis of power consumption at the component-level, node-level, and system-level, from all 4,626 Summit compute nodes, each with over 100 metrics at 1Hz frequency over the entire year of 2020. We also investigate the power characteristics and energy efficiency of over 840k Summit jobs and 250k GPU failure logs for further operational insights. To the best of our knowledge, this is the first systematic analysis of power data of HPC system at this scale.

As we move into the exascale era, supercomputers grow larger, denser, more heterogeneous, and ever more complex. Operating such machines reliably and efficiently requires deep insight into the operational parameters of the machine itself as well as its supporting infrastructure. To fulfill this need, early adopter sites have started the development and deployment of Operational Data Analytics (ODA) frameworks allowing the continuous monitoring, archiving, and analysis of near realtime performance data from the machine and infrastructure levels, providing immediately actionable information for multiple operational uses. To understand their ODA goals, requirements, and use cases, we have conducted a survey among eight early adopter sites from the US, Europe, and Japan that operate top 50 high-performance computing systems. We have assessed the technologies leveraged to build their ODA frameworks, identified use cases and other push and pull factors that drive the sites' ODA activities, and report on their operational lessons.

We present the design and implementation of Data Jockey, a data management system for HPC multi-tiered storage systems. As a centralized data management control plane, Data Jockey automates bulk data movement and placement for scientific workflows and integrates into existing HPC storage infrastructures. Data Jockey simplifies data management by eliminating human effort in programming complex data movements, laying datasets across multiple storage tiers when supporting complex workflows, which in turn increases the usability of multitiered storage systems emerging in modern HPC data centers. Specifically, Data Jockey presents a new data management scheme called “goal driven data management” that can automatically infer low-level bulk data movement plans from declarative high-level goal statements that come from the lifetime of iterative runs of scientific workflows. While doing so, Data Jockey aims to minimize data wait times by taking responsibility for datasets that are unused or to be used, and aggressively utilizing the capacity of the upper, higher performant storage tiers. We evaluated a prototype implementation of Data Jockey under a synthetic workload based on a year's worth of Oak Ridge Leadership Computing Facility's (OLCF) operational logs. Our evaluations suggest that Data Jockey leads to higher utilization of the upper storage tiers while minimizing the programming effort of data movement compared to human involved, per-domain adhoc data management scripts.

Current key value stores rely on DRAM based inmemory architectures where scalability is limited by high power and low density of DRAM. As an alternative, flash SSDs has been explored because of the merits of low power, high density and high internal parallelism. However, the unpredictable latency caused by SSD internal resource conflicts challenges the use of flash SSDs. To address this issue, we present HIOPS-KV a storage I/O software stack for key value stores. HIOPS-KV exploits multiple solid-state drives (SSDs) to control the latencies. With replicas, HIOPS-KV avoids structural collisions which cause long latency operations by spreading colliding operations to distinct devices. For evaluation, we integrated HIOPS-KV into memcached on a low cost high IOPS SSD system built with PC components. At 32 YCSB clients, our system was capable of 117k ops/sec with 263 us average latency showing approximately 4ms at the 99th percentile latency.

Fast storage devices are an emerging solution to satisfy data-intensive applications. They provide high transaction rates for DBMS, low response times for Web servers, instant on-demand paging for applications with large memory footprints, and many similar advantages for performance-hungry applications. In spite of the benefits promised by fast hardware, modern operating systems are not yet structured to take advantage of the hardware’s full potential. The software overhead caused by an OS, negligible in the past, adversely impacts application performance, lessening the advantage of using such hardware. Our analysis demonstrates that the overheads from the traditional storage-stack design are significant and cannot easily be overcome without modifying the hardware interface and adding new capabilities to the operating system. In this article, we propose six optimizations that enable an OS to fully exploit the performance characteristics of fast storage devices. With the support of new hardware interfaces, our optimizations minimize per-request latency by streamlining the I/O path and amortize per-request latency by maximizing parallelism inside the device. We demonstrate the impact on application performance through well-known storage benchmarks run against a Linux kernel with a customized SSD. We find that eliminating context switches in the I/O path decreases the software overhead of an I/O request from 20 microseconds to 5 microseconds and a new request merge scheme called Temporal Merge enables the OS to achieve 87% to 100% of peak device performance, regardless of request access patterns or types. Although the performance improvement by these optimizations on a standard SATA-based SSD is marginal (because of its limited interface and relatively high response times), our sensitivity analysis suggests that future SSDs with lower response times will benefit from these changes. The effectiveness of our optimizations encourages discussion between the OS community and storage vendors about future device interfaces for fast storage devices.

Over the past decades, caching has become the key technology used for bridging the performance gap across memory hierarchies via temporal or spatial localities; in particular, the effect is prominent in disk storage systems. Applications that involve heavy I/O activities, which are common in the cloud, probably benefit the most from caching. The use of local volatile memory as cache might be a natural alternative, but many well-known restrictions, such as capacity and the utilization of host machines, hinder its effective use. In addition to technical challenges, providing cache services in clouds encounters a major practical issue (quality of service or service level agreement issue) of pricing. Currently, (public) cloud users are limited to a small set of uniform and coarse-grained service offerings, such as High-Memory and High-CPU in Amazon EC2. In this paper, we present the cache as a service (CaaS) model as an optional service to typical infrastructure service offerings. Specifically, the cloud provider sets aside a large pool of memory that can be dynamically partitioned and allocated to standard infrastructure services as disk cache. We first investigate the feasibility of providing CaaS with the proof-of-concept elastic cache system (using dedicated remote memory servers) built and validated on the actual system, and practical benefits of CaaS for both users and providers (i.e., performance and profit, respectively) are thoroughly studied with a novel pricing scheme. Our CaaS model helps to leverage the cloud economy greatly in that 1) the extra user cost for I/O performance gain is minimal if ever exists, and 2) the provider's profit increases due to improvements in server consolidation resulting from that performance gain. Through extensive experiments with eight resource allocation strategies, we demonstrate that our CaaS model can be a promising cost-efficient solution for both users and providers.