Challenges for Implementing FAIR Digital Objects with High Performance Workflows

New types of workflows are being used in science that couple traditional distributed and high-performance computing (HPC) with data-intensive approaches, orchestrate ensembles numerical simulations artificial intelligence (AI) models. Such may use AI models to supplement computation where be too computationally expensive, automate trivial yet time consuming operations, perform preliminary selections among intractable numbers combinations domains as diverse protein binding, fine-grid climate simulations, drug discovery. They offer renewed opportunities for scientific research but exhibit high computational, storage communications requirements [Goble et al. 2020, Al-Saadi 2021, da Silva 2021]. These can orchestrated by workflow management systems (WMS) built upon composable blocks facilitate task placement resource allocation parallel executions on performance [Lee Merzky The communities running these kinds have been slow adopt Findable, Accessible, Interpretable, Re-usable (FAIR) principles, part due the complexity life cycles, numerous WMS, specificity HPC rapidly evolving architectures software stacks, execution modes require managers batch schedulers [Plale FAIR Digital Objects (FDO) encapsulate bit sequences data, metadata, persistent identifiers (PID) help promote adoption FAIR, enable knowledge extraction dissemination, contribute re-use [De Smedt 2020]. As typically data during planning execution, FDOs particularly adapted [Wittenburg But benefits such automating processing actionable DO collections cannot realized without main components rich metadata clear identifiers, universally adopted community. still elusive digital objects. Some added after results produced, not described controlled vocabularies, left unconstrained, resulting inefficient processes loss knowledge. Persistent at publication supporting conclusions, so only a very small amount shared outside community researchers “in know”. In this conceptual work, one distinguish several present both common specific challenges development canonical infrastructure implementation FDO we discuss below: result represent computational obtained when program complete, contain measures from code optimization parallel, heterogeneous architectures, intermediate states checkpointing. All should include environment system specifications which was executed enough re-usability [Pouchard 2019]. Containers often capture dependencies between underlying libraries versions installation [Lofstead 2015, Olaya containers published repositories made available registered resolvers. For instance, attribute Object Identifier github, must additional step registering into Zenodo. extracted context framework including will attribution linking workflow. Computational machine learning predictions form stochastic training non-deterministic Neural networks deep related provenance selection quantities needed an results. What information needs included encapsulating make it re-usable? description method, experiment recommended [Gundersen Kjensmo 2018] instantiated collection. To re-usable, model architecture, platform its version, submission script contains hyperparameters, function, size number epochs Challenges objects containing those size, reduction. Performance scale tends large, thus principled approach is determine counters reproducibility application [Patki variables selected show their impact methods selection: do outliers metrics? thresholds qualify outliers, what overall execution? A key contributor failure important “bolted on” fact piecemeal, cumbersome, manner impedes further analysis. An appropriate level abstraction needed. Capturing automatically take account granularity across layers levels. Intermediate fuse multiple sources stages [Nicolae 2022]. tools already exist. Darshan scalable tool summarizing Input/Output file characteristics [Dai 2019], Radical Cybertools [Merzky 2021] produce graph execution. could they path forward services would guarantee encapsulation DOs favorable re-use.