Please, forgive the silly joke in the title of this semi-serious post, but lately I have been thinking about the strange fate of an area of general computing that I have spent more and more time in recently, as in the near and far past. For my job, I have utilized a series of scientific HPC clusters worldwide to solve multiple computing problems most efficiently by distributing computation across numerous nodes. Over the last thirty years, all such platforms have consistently shared the same common characteristics, which invariably pose a problem in their use for the average scientist (often a young/junior dedicated to a short-term project) in any application domain.
To use Fred Brooks' definition, HPC technologies have both intrinsic and incidental fallacies for such users category. The intrinsic one is due to the inner complexity of creating a parallel and distributed solution to any problem, possibly in a way that does not harm the final implementation due to the increase in communication time among computational agents. This is already a relevant problem per se, which can often be out of the abilities, knowledge, and interests of the average researcher in bioinformatics, physics, mathematics, remote sensing, or whatever other research domain.
The incidental fallacy is instead always due to the accessibility of platforms and the technologies used for their implementation. At large, all such HPC clusters are a large pool of multi-core hosts with plenty of memory and connected with multiple high-speed networks for implementing some sort of multi-tier distributed POSIX file system and/or object storage. Users can log in on a limited number of such hosts that are connected to all others and run some type of scheduling system (e.g., Slurm or HTcondor) where multiple computational nodes can be reserved for a limited period of time to execute batch jobs or even an interactive one (mainly for debugging). In most cases, such clusters can also be used with some MPI/OpenMP implementations for proper parallel computational modeling based on message passing among computing agents that run on multiple cores and hosts, with or without multi-threading. Alternatively, GPUs can also be reserved and exploited via Cuda/OpenCL. In many cases, such implementations are vendor-oriented and trigger the need to adopt specific libraries and compilers that add another layer of complexity to implementations.
The incidental problems start when the casual users discover that all such computing nodes invariably run some legacy enterprise Linux distribution that is maintained for a period of ten years or even more, until a full reinstallation of the whole cluster. On top of such legacy systems (that are for any practical use simply unusable as such) these scientific clusters give essentially a few different mechanisms for creating a general computational environment:
- Environment Modules
- Containers (Singularity or Apptainer)
- Anaconda/Miniconda-like environment (or free forks like Miniforge)
- Some specific software/application to run
But for containers, the other solutions are all binary-based hubs, which could expose them to possible breakages when the application developed needs to access exotic language bindings for extensions, and the poor users enter the mysterious and dangerous world of ABI violations and a chain of broken dependencies. Even, often, such hubs are not always consistent, and any upgrade by the admin team exposes them to sudden breakages from night to day.
The final solution (or apparently so) nowadays is using containerization and a target environment where the user code can find all and only the correct dependencies and versions for the whole software stack of the application. This, at least, until the third-party hubs of base distributions and languages ensure complete consistency and retain past binaries and versions for any medium/long-term need. Of course, a full source-based stack with proper version tracking a la Guix would help to avoid dependencies on external binary hubs and seems the way to go. Indeed, a small group of interest in such a solution has existed for a few years, but I am unaware of so many HPC clusters that consistently propose this kind of implementation for users. That said, writing Guile Scheme descriptors for preparing an execution environment may not be within the reach of the average researcher in biochemistry or astrophysics.
Unfortunately, as I wrote in a past post on this digital site, this moves the whole responsibility of a software stack maintenance onto the shoulders of the final users, who are often the infamous junior profiles I mentioned before. These are non-IT specialists who should adopt such HPC platform to implement solutions as part of their daily job in their special scientific domain.
The result, to be honest, is that the average researcher simply tries to avoid the whole thing as soon as possible because of the significant complexity that the entire thing involves, while the private sector introduced specialistic roles of data and software engineers to manage such problems properly (which is the only reasonable approach, indeed). Adding insult to injury, in some academic areas, such interests in HPC are also viewed with contempt or as a waste of time, if not openly discouraged.
All this explains why a roundabout in any of the significant HPC clusters worldwide often guarantees hilarious experiences in terms of who is doing what and how.
Sometimes, I almost feel like I can hear them swearing...
For comments, join the discussion on Mastodon.