Hacker News Reader

Comment by KenoFischer | Hacker News Reader

Counting the number of galaxies in these image datasets is actually a highly non-trivial, supercomputer-scale problem. I worked on this a few years back! Paper link for those interested: https://arxiv.org/pdf/1801.10277.pdf

View on HN

I am constantly amazed by the people that participate in HN. Everything I do seems so trivial compared to what some of you guys do.

Well, you're not alone.

Why supercomputer? Isn't this kind of classification task embarrassingly parallelizable?

Modern supercomputers are obscenely parallel machines built to chew through embarrassingly parallelizable tasks.

So is a GPU

Might depend on the level of embarrassment.

Interesting that they did this with Julia, with 83% of instructions being AVX-512 (if I'm reading it correctly).

Does anyone know if Julia's GPU capabilities could have been leveraged on say a cluster of NVIDIA A100/V100?

This work is four years old (with development happening before that), so the Julia GPU capabilities probably weren't good enough at the time. If you wanted to do it today, that'd probably be the way to do it, but would need some benchmarking.

A lot of modern supercomputer use/have GPUs. But most GPUs had very bad fp64 compute capabilities, so they were not really used for anything requiring precision for a long time.

Forgive me if I'm wrong but I thought supercomputer time was usually not allocated to embarrassingly parallel tasks. While they certainly can do those tasks well they're a waste of a distributed system with expensive, high bandwidth fiber connections between nodes.

When I spent (a relatively small amount of) time working with one this was the main thing the director drilled into my head. Use it to solve large, parallel problems that require lots of intranode communication of intermediate results. Embarrassingly parallel problems can be solved on cheaper hardware like GPUs.

When the data no longer fits inside a compute resource (a node, or even a rack), you are by necessity going to be distributed. Communication is a fact of life when the problem-size grows. This is true also for GPU-based computing.

It's not entirely embarrassingly parallel because photons from nearby light sources interact, so you need to do a parameter exchange. That said, it is true that it doesn't need the low latency fabric that supercomputers traditionally used (though these days the architecture isn't all that different from what you'd have in a traditional data center). We were actually approached by Microsoft to duplicate the demo on Azure, but they ended up being unable to find enough spare capacity for us to run it at full scale, so we showed off a small scale demo. In this particular case, the supercomputing center was quite interested in our application because they wanted to test their setup with more "data analysis heavy" workloads, since they had invested a lot in SSD caches for this kind of work, but most of their workloads didn't require it.

Congratulations! Thank you for making the world better with such amazing work!

Not the world, but the universe!

We can only improve the subset of the universe that lies within our future light cone, and even then only with a significantly better space program.

Reply on news.ycombinator.com