Hacker News Reader

Comment by Retr0id | Hacker News Reader

It seems like they're targeting low-end systems (384MB RAM for the desktop version), but also talk about supercomputers? If they can pull that off, it would make it uniquely versatile. Musl is very lightweight, but it's also not especially fast (for example, compare allocator benchmarks). But maybe supercomputer workloads don't really care about libc performance in the first place? I'd be curious to know what aspects they think make this distro suited to supercomputing.

View on HN

Our aim is to offer a non-glibc Linux platform for supercomputing environments. This involves a tremendous amount of work to be taken seriously and we are currently in the earliest stage of this. Follow our blog (https://blog.adelielinux.org/) for updates in this space.

What are the benefits of a non-glibc supercomputing environment?

Diversity in the ecosystem, possible reduction in memory footprint, ability to discover and fix issues in upstream projects that musl-based distros don't currently package, motivation to support CUDA (and other) runtimes that explicitly target glibc, unlock possible sources of funding to pay developers to work on these areas.

The ability to audit a smaller codebase could be compelling for some sensitive environments.

There is not an urgent or immediate problem with glibc in the HPC space, per se.

Do you have any benchmarks etc. that compare performance of HPC workloads on glibc vs musl?

In progress. We have access to an ARM-based supercomputer and are actively collaborating with scientists who have access to x86_64 systems. It will be a while before we have results to publish. We need to be diligent in our testing methodologies.

Have you published your testing methodologies?

You know the quickest way to not be taken seriously? Lying on your website.

https://news.ycombinator.com/item?id=38891721#38893535

I dunno if allocation speed is that critical to HPC... presumably in that environment, you're allocating big arrays up front and then reusing them. Maybe someone can set me straight on this?

It's been a while since I've seen HPC code, but there's more people who can write average code that need HPC than there are specialists to do it well.

I could see some people being interested in building compute clusters where the individual nodes are cheap, low-memory devices. (Think Raspberry Pi clusters.)

That does sound like a pretty niche thing, though.

Reply on news.ycombinator.com