Strategic Benchmarking and Acquisition of Next-Generation Supercomputers at ExxonMobil
Abstract: In the realm of high-performance computing, supercomputers are tailored for specific tasks, and not every machine can address every problem. A supercomputer optimized for one domain may face challenges when tackling problems from a different field. Selecting the next generation of supercomputers requires a meticulous benchmarking process to ensure they meet the desired performance and efficiency standards. This presentation delves into ExxonMobil’s strategic approach to acquiring new supercomputers.
Given the proprietary nature of our code, sharing it with vendors is not an option. Our code base contains hundreds of thousands of lines of code, making it impractical to compile and run on every available hardware option within the timeframe for making a purchase decision. Additionally, standard industry benchmarks might not capture the true complexities and performance characteristics of our applications. Instead, we develop synthetic applications that simulate the computational and I/O complexities of our actual workloads by identifying and mimicking bottlenecks. Understanding which parts of the code are performance-critical is essential in developing these synthetic benchmarks. These benchmarks evolve over time to reflect shifting bottlenecks, providing a reliable means to evaluate potential hardware.
Power consumption is another critical factor. Our data center operates at maximum power capacity, necessitating the decommissioning of older machines when new ones are acquired. Additionally, the timeline for hardware availability is a significant consideration. A top-tier machine arriving a year later may be less beneficial than a slightly less powerful one available within months. Understanding the performance gains and economic implications of these decisions is essential. ExxonMobil’s decision-making process for purchasing new supercomputers involves a comprehensive analysis of benchmarking results, costs, power constraints, and hardware timelines. This thorough approach ensures that we select supercomputers that align with our computational needs and operational constraints, ultimately enhancing our performance and efficiency.
Bio: Karthik Neerala Suresh is a leading figure in high-performance computing (HPC) for seismic imaging, blending deep technical expertise with a unique philosophy on computational accuracy. With six years of intensive experience at ExxonMobil, he has built a reputation as a developer who not only understands the complexities of HPC but also drives innovation within the field. Karthik’s academic journey in Computational Mechanics, with a specialization in numerical analysis and high-performance computing from Swansea University and UPC Barcelona, laid a robust foundation for his career. His master’s degree equipped him with the skills to tackle intricate numerical challenges and leverage HPC capabilities in seismic data analysis. Professionally, Karthik has been a vital member of ExxonMobil’s team, benchmarking hardware to guide business purchase decisions. He played a key role in developing synthetic applications that accurately represent ExxonMobil’s workloads, which have been instrumental in multiple procurement processes. A strong advocate for data-driven decision-making, Karthik works tirelessly to gather information that mitigates risks in supercomputer purchases. Karthik has an extensive track record in optimizing, maintaining, and porting reverse time migration and full waveform inversion applications for both CPUs and GPUs. His hands-on experience in fine-tuning these critical seismic imaging applications demonstrates his commitment to enhancing performance while ensuring computational accuracy.
Authors: Karthik Neerala Suresh (ExxonMobil), Kirsten Byers (ExxonMobil), Rahul Sampath (ExxonMobil) and Jordan Pomeroy (ExxonMobil)
