Power consumption and associated carbon emissions are increasingly critical challenges for large-scale computing. Recent research proposes exploiting stranded power - uneconomic renewable power - for green supercomputing in a system called Zero-Carbon Cloud (ZCCloud) [1], [2], [3]. These efforts studied production supercomputing workloads on stranded-power based computing resources, demonstrating their achievable productivity. We explore economic viability of stranded-power based supercomputing, using three datacenter total-cost-of-ownership (TCO) models to study cost-effectiveness. These studies show that ZCCloud's approach can be cost-effective in the USA today, and is even more attractive in regions with higher power prices (e.g., Japan, Germany), achieving cost advantages as large as 50 percent. Environmental and power-grid benefits are a further advantage. We also explore the sensitivity of these results to changes in hardware TCO; cheaper hardware or longer lifetimes magnify the attractiveness of stranded-power based approaches, yielding advantages as large as 91 percent. These results are robust across different TCO models. Finally, we study extreme-scale supercomputers (>100 MW), finding stranded-power can increase peak capability per cost by as much as 80 percent.

Traditional cloud stacks are designed to tolerate random, small-scale failures, and can successfully deliver highly-available cloud services and interactive services to end users. However, they fail to survive large-scale disruptions that are caused by major power outage, cyber-attack, or region/zone failures. Such changes trigger cascading failures and significant service outages. We propose to understand the reasons for these failures, and create reliable data services that can efficiently and robustly tolerate such large-scale resource changes.

In supercomputer centers, available power, cooling, or carbon footprint often limits supercomputer performance. We propose a new approach to continue scaling that avoids many of these limits, augmenting a traditional system with another that employs only "wasted" renewable power, stranded power. This excess power cannot be economically distributed through grid, and is only intermittently available. We call this approach Zero-carbon Cloud (ZCCloud). We explore the potential benefits of unreliable resources with production DOE HPC workloads using a simple periodic model, and identify job types that benefit most (capability jobs and on-time jobs). The benefits scale with duty factor and resource quantity. Next, to create realistic models of "stranded power" we study 28 months of Mid-continent Independent System Operator (MISO) power market history (1,259 generators, 77 million 5-minute intervals). We find that opportunity varies, but the best single wind site can provide 80% duty factor, and 20MW average stranded power. Combining sites further improves duty factor. With resource volatility models from the MISO study, we simulate production DOE HPC workloads and find that stranded power HPC, ZCCloud, can provide significant benefit, decreasing average job-wait time by 50%.

Graph processing is important for a growing range of applications. Current performance studies of parallel graph computation employ a large variety of algorithms and graphs. To explore their robustness, we characterize behavior variation across algorithms and graph structures at different scales. Our results show that graph computation behaviors, with up to 1000-fold variation, form a very broad space. Any inefficient exploration of this space may lead to narrow understanding and ad-hoc studies. Hence, we consider constructing an ensemble of graph computations, or graph-algorithm pairs, to most effectively explore this graph computation behavior space. We study different ensembles of parallel graph computations, and define two metrics to quantify how efficiently and completely an ensemble explores the space. Our results show that: (1) experiments limited to a single algorithm or a single graph may unfairly characterize a graph-processing system, (2) benchmarks exploring both algorithm and graph diversity can significantly improve the quality (30% more complete and 200% more efficient), but must be carefully chosen, (3) some algorithms are more useful than others in benchmarking, and (4) we can reduce the complexity (number of algorithms, graphs, runtime) while conserving the benchmarking quality.