Search

This presentation covers some of the highlights or the struggles, successes, and innovations that occurred in the process of standing up and running the Summit system at Oak Ridge National Laboratory.

The Cray XK7 Titan was the top supercomputer system in the world for a long time and remained critically important throughout its nearly seven year life. It was an interesting machine from a reliability viewpoint as most of its power came from 18,688 GPUs whose operation was forced to execute three rework cycles, two on the GPU mechanical assembly and one on the GPU circuitboards. We write about the last rework cycle and a reliability analysis of over 100,000 years of GPU lifetimes during Titan’s 6-year-long productive period. Using time between failures analysis and statistical survival analysis techniques, we find that GPU reliability is dependent on heat dissipation to an extent that strongly correlates with detailed nuances of the cooling architecture and job scheduling. We describe the history, data collection, cleaning, and analysis and give recommendations for future supercomputing systems. We make the data and our analysis codes publicly available.

Supercomputers stand as a fundamental tool for developing our understanding of the universe. State-of-the-art scientific simulations, big data analyses, and machine learning executions require high performance computing platforms. Such infrastructures have been growing lately with the addition of thousands of newly designed components, calling their resiliency into question. It is crucial to solidify our knowledge on the way supercomputers fail. Other recent studies have highlighted the importance of characterizing failures on supercomputers. This paper aims at modelling component failures of a supercomputer based on Mixed Weibull distributions. The model is built using a real-life multi-year failure record from a leadership-class supercomputer. Using several key observations from the data, we designed an analytical model that is robust enough to represent each of the main components of supercomputers, yet it is flexible enough to alter the composition of the machine and be able to predict resilience of future or hypothetical systems.

Extreme-scale computing systems are required to solve some of the grand challenges in science and technology. From astrophysics to molecular biology, supercomputers are an essential tool to accelerate scientific discovery. However, large computing systems are prone to failures due to their complexity. It is crucial to develop an understanding of how these systems fail to design reliable supercomputing platforms for the future. This paper examines a five-year failure and workload record of a leadership-class supercomputer. To the best of our knowledge, five years represents the vast majority of the lifespan of a supercomputer. This is the first time such analysis is performed on a top 10 modern supercomputer. We performed a failure categorization and found out that: i) most errors are GPUrelated, with roughly 37% of them being double-bit errors on the cards; ii) failures are not evenly spread across the physical machine, with room temperature presumably playing a major role; and iii) software errors of the system bring down several nodes concurrently. Our failure rate analysis unveils that: i) the system consistently degrades, being at least twice as reliable at the beginning, compared to the end of the period; ii) Weibull distribution closely fits the mean-time-between-failure data; and iii) hardware and software errors show a markedly different pattern. Finally, we correlated failure and workload records to reveal that: i) failure and workload records are weakly correlated, except for certain types of failures when segmented by the hours of the day; ii) several categories of failures make jobs crash within the first minutes of execution; and iii) a significant fraction of failed jobs exhaust the requested time with a disregard of when the failure occurred during execution. Index Terms-Fault tolerance, resilience, failure analysis, high performance computing.

Summit, the latest flagship supercomputer deployed at Oak Ridge Leadership Computing Facility (OLCF), became the number one system in the TOP500 [17] list in June 2018 and retained its top spot in the November 2018 list. An extensive acceptance test plan was developed to evaluate the unique features introduced in the Summit architecture and system software stack. The acceptance test also includes tests to ensure that the system is reliable, stable, and performant.

In 2015, OLCF's Titan supercomputer experienced a significant increase in GPU related job failures. The impact on jobs was serious and OLCF decided to replace ~50% of the GPUs. Unfortunately, jobs using more than 20% of the machine (i.e., leadership jobs) continued to encounter higher levels of application failures. These jobs contained significant amounts of both the low-failure rate and high-failure rate GPUs. The impacts of these failures are more adversely felt by leadership jobs due to longer wait times, runtimes, and higher charge rates. In this work, we have designed techniques to increase the use of low-failure GPUs in leadership jobs through targeted resource allocation. We have employed two complementary techniques, updating both the system ordering and the allocation mechanisms. Using simulation, the application of these techniques resulted in a 33% increase in low-failure GPU hours being assigned to leadership jobs. Our GPU Age-Aware Scheduling has been used in production on Titan since July of 2017.

CORAL, the Collaboration of Oak Ridge, Argonne and Livermore, is fielding two similar IBM systems, Summit and Sierra, with NVIDIA GPUs that will replace the existing Titan and Sequoia systems. Summit and Sierra are currently ranked No. 1 and No. 3, respectively on the Top500 list. We discuss the design and key differences of the systems. Our evaluation of the systems highlights the following. Applications that fit in HBM see the most benefit and may prefer more GPUs; however, for some applications, the CPU-GPU bandwidth is more important than the number of GPUs. The node-local burst buffer scales linearly, and can achieve a 4X improvement over the parallel file system for large jobs; smaller jobs, however, may benefit from writing directly to the PFS. Finally, several CPU, network and memory bound analytics and GPU-bound deep learning codes achieve up to a 11X and 79X speedup/node, respectively over Titan.

In preparation for Summit, Oak Ridge National Laboratory’s next generation supercomputer, two IBM Power-based systems were deployed in late 2016 at the Oak Ridge Leadership Computing Facility (OLCF). This paper presents a detailed description of the acceptance of the first IBM Power-based early access systems installed at the OLCF. The two systems, Summitdev and Tundra, contain IBM POWER8+ processors with NVIDIA Pascal GPUs and were acquired to provide researchers with a platform to optimize codes for the Power architecture. In addition, this paper presents early functional and performance results obtained on Summitdev with the latest software stack available.

The high computational capability of graphics processing units (GPUs) is enabling and driving the scientific discovery process at large-scale. The world's second fastest supercomputer for open science, Titan, has more than 18,000 GPUs that computational scientists use to perform scientific simulations and data analysis. Understanding of GPU reliability characteristics, however, is still in its nascent stage since GPUs have only recently been deployed at large-scale. This paper presents a detailed study of GPU errors and their impact on system operations and applications, describing experiences with the 18,688 GPUs on the Titan supercomputer as well as lessons learned in the process of efficient operation of GPUs at scale. These experiences are helpful to HPC sites which already have large-scale GPU clusters or plan to deploy GPUs in the future.

As we approach exascale, the scientific simulations are expected to experience more interruptions due to increased system failures. Designing better HPC resilience techniques requires understanding the key characteristics of system failures on these systems. While temporal properties of system failures on HPC systems have been well-investigated, there is limited understanding about the spatial characteristics of system failures and its impact on the resilience mechanisms. Therefore, we examine the spatial characteristics and behavior of system failures. We investigate the interaction between spatial and temporal characteristics of failures and its implications for system operations and resilience mechanisms on large-scale HPC systems. We show that system failures have "spatial locality" at different granularity in the system, study impact of different failure-types, and investigate the correlation among different failure-types. Finally, we propose a novel scheme that exploits the spatial locality in failures to improve application and system performance. Our evaluation shows that the proposed scheme significantly improves the system performance in a dynamic and production-level HPC system.

Increase in graphics hardware performance and improvements in programmability has enabled GPUs to evolve from a graphics-specific accelerator to a general-purpose computing device. Titan, the world's second fastest supercomputer for open science in 2014, consists of more dum 18,000 GPUs that scientists from various domains such as astrophysics, fusion, climate, and combustion use routinely to run large-scale simulations. Unfortunately, while the performance efficiency of GPUs is well understood, their resilience characteristics in a large-scale computing system have not been fully evaluated. We present a detailed study to provide a thorough understanding of GPU errors on a large-scale GPU-enabled system. Our data was collected from the Titan supercomputer at the Oak Ridge Leadership Computing Facility and a GPU cluster at the Los Alamos National Laboratory. We also present results from our extensive neutron-beam tests, conducted at Los Alamos Neutron Science Center (LANSCE) and at ISIS (Rutherford Appleron Laboratories, UK), to measure the resilience of different generations of GPUs. We present several findings from our field data and neutron-beam experiments, and discuss the implications of our results for future GPU architects, current and future HPC computing facilities, and researchers focusing on GPU resilience.

The metric, Power Usage Effectiveness (PUE), has been successful in improving energy efficiency of data centers, but it is not perfect. One challenge is that PUE does not account for the power distribution and cooling losses inside IT equipment. This is particularly problematic in the HPC (high performance computing) space where system suppliers are moving cooling and power subsystems into or out of the cluster. This paper proposes two new metrics: ITUE (IT-power usage effectiveness), similar to PUE but “inside” the IT and TUE (total-power usage effectiveness), which combines the two for a total efficiency picture. We conclude with a demonstration of the method, and a case study of measurements at ORNL’s Jaguar system. TUE provides a ratio of total energy, (internal and external support energy uses) and the specific energy used in the HPC. TUE can also be a means for comparing HPC site to HPC site.

According to the U.S. Energy Information Administration, LNG is projected to become a much larger share of U.S. natural gas consumption, rising from current levels of around 2.5% of total natural gas consumption to 12.4% by 2030. Because natural gas and LNG are substitutes, natural gas prices are expected to be an important determinant of LNG imports. Furthermore, an increasing share of LNG is traded under short-term contracts with spot shipments being diverted to markets offering the highest returns (netbacks). Relative natural gas prices as well as LNG transportation costs are important determinants of LNG netbacks. This paper examines the empirical relationship between U.S. LNG imports, the Henry Hub price of natural gas relative to U.K. and Asia gas prices, and a proxy for LNG transportation costs using monthly data from 1997 to 2007. Granger causality tests, error variance decomposition, and impulse response analyses using a VAR model are employed to establish Granger-causality as well as the dynamics of natural gas prices and LNG transportation innovations on LNG imports.