HPC Kitchen metaphor

This is a series of videos which compares HPC and more broadly scientific computing with cooking or other things in real life. The goal is to make approaching computing easier and

Warning

This page is under construction and not all descriptions/transcripts are here yet. Currently, YouTube has the videos and full commentary, but I hope to make that available without YouTube.

[intro]: computing vs cooking

https://www.youtube.com/watch?v=yqGtnA7CUtU&list=PLZLVmS9rf3nNDHRo1Baz_JVQWDI0mTYyB

Advanced computing resources like HPC (high-performance computing) clusters may seem intimidating, but really all the pieces are the same as a normal computer. The difference is you have to be able to use and coordinate all the resources. This sets you an the path by introducing you to a metaphor relating computing to cooking.

[data storage]: Understanding how data is stored and moved

https://www.youtube.com/watch?v=JAR9xyy5rcE&list=PLZLVmS9rf3nNDHRo1Baz_JVQWDI0mTYyB

It’s all about data these days, but where does it all go? How do understand what we have? We compare storage in a HPC cluster (or really, anywhere) to storage of food in a kitchen. We’ll see how the hierarchy of size vs speed goes and how we access data on different servers. This doesn’t teach details about storage but prepares you to learn details in a future course.

[storage-performance]: Understanding the speed of data access and transfer

https://www.youtube.com/watch?v=9siGLV8pZ5A&list=PLZLVmS9rf3nNDHRo1Baz_JVQWDI0mTYyB

When using supercomputers, data movement can actually be what slows you down surprisingly often. A modern GPU can easily pull in data faster than it can be delivered from storage, if you don’t preprocess it well. In this video, we go over the basic aspects of storage performance… with food.

[parallel]: methods to run on multiple processors

https://www.youtube.com/watch?v=I6fBq9HN3P4&list=PLZLVmS9rf3nNDHRo1Baz_JVQWDI0mTYyB

Running programs in parallel: that’s what everyone thinks they want to use a cluster for. But there are actually different ways this can work, and you need to be able to distinguish between them, so that you can run them properly. This broadly explains what these methods are, so that you can understand later technical documentation.

[slurm]: the Slurm job scheduler spreads tasks to the cluster

https://www.youtube.com/watch?v=Y73A7lXISxU&list=PLZLVmS9rf3nNDHRo1Baz_JVQWDI0mTYyB

We have talked about running stuff in parallel, but how does it get connected to hardware (processors, memory)? If we have one cluster for everyone, how do people share? That’s what this is all about.

[containers/environments]: Moving code around

https://www.youtube.com/watch?v=Ag7wcAwU_Jw&list=PLZLVmS9rf3nNDHRo1Baz_JVQWDI0mTYyB

When you go to a different kitchen to cook, everything is different and it slows you down. This is a worse problem for computers, which don’t have a brain and thus the recipe has to be exact. As you get farther and farther into scientific coding, you’ll see how hard of a difficulty this “installing software and makin g it work” is. Environments and containers solve this problem.

[how-to-learn]: Don’t give up in learning computing

https://www.youtube.com/watch?v=evJV-02poDc&list=PLZLVmS9rf3nNDHRo1Baz_JVQWDI0mTYyB

This is more of a fireside chat, to encourage you not to give up in learning computing. It can often feel there is so much to learn and it’s almost impossible. It can also feel like you can’t do it yourself. Well, most people don’t - all learning is somehow collaborative. Take the time to work together to share skills. Take the time to follow others even in practical computing skills, not just academic. And help others.

[data-management]: Data management is as important as data storage

https://www.youtube.com/watch?v=Q5A7n7mu-AI&list=PLZLVmS9rf3nNDHRo1Baz_JVQWDI0mTYyB

People may think that getting a big enough storage space for your data is all you need. There’s far more than that. It’s easy for data (and kitchens) to become a huge mess without some care. It’s even worse when you work together. Unfortunately, it’s far too common for people to be too rushed to organize their stuff well, leading to worse problems down the line. Take care of your data and you data will :strikeout:`take care of` be valuable to you.

[big-vs-small-jobs]: The pitfalls of big jobs

https://www.youtube.com/watch?v=Qrql8rGfRVo&list=PLZLVmS9rf3nNDHRo1Baz_JVQWDI0mTYyB

I was eating at a restaurant and their group policy said large groups have a different menu. Why is that? Larger groups are a less efficient use of tables, because of communication and synchronization overhead. In the same way, larger jobs on the cluster can be less efficiency because of overheads.

[reading-docs]: Making sense of information overload

https://www.youtube.com/watch?v=E9w-MNaXkDw&list=PLZLVmS9rf3nNDHRo1Baz_JVQWDI0mTYyB

Have you ever been faced with on overwhelming amount of information you need to go through before you can start something? Often times, it’s because there is actually so much information, but there are strategies for dealing with it. Let me talk about how I approach information overload…

Transcript

Hello, and welcome to another low-quality HPC Kitchen video. This time we’re talking about documentation, or generally finding information.

So, my partner’s just gotten back from traveling, and this time instead of going through the checklist myself, I printed it out and gave it to them and they looked at this and said “this is too long, I can’t go through it all before traveling and well I mean partly yes that’s true because this [sheet] has everything that I’ve ever needed to think about before traveling. For the last 20 years or so, anytime I’ve forgotten something or had to do something before traveling, I would add it to the list, and it becomes long.

So what does this mean? Is longer worse? So in some sense you might say so because it’s too hard to read. But the point of the checklist is to have everything that you might need on it, so you can check through it, and once you’ve read it you make sure you’ve got everything. And then you can go off and be confident you haven’t forgotten anything. If it had less on it it would be easier to go through, but then you would still have to go think “what am I missing?”, and then think of that, and maybe forget things. So it’s actually longer than it would need otherwise. So actually the checklist needs to be long in order to be useful.

So what’s the metaphor here? The metaphor is the documentation to use all our computing resources. So if you try to print it out it would be around this long. Of course I didn’t actually print it out just for a video. I’m not crazy like that [I’m crazy in other ways, like thinking of this topic]. I printed out the first page and then the rest is some other paper I’ve gotten. So this is long. What do we do? So, do we want it to be shorter or longer?

So if you ask me the answer here is organization. So if you look at my packing list it’s sorted into 13 categories with titles that tell you when you may need each of the different things in it. So for example right before you walk out the door you look at the “final check” section. When you’re packing you would look at the “clothes” and maybe “information” section. If no one’s staying at home and you need to turn stuff off and pack stuff up before leaving, there’s a “if no one is staying section”, and so on. So this allows you to sort of have something long, but also go through and be able to use it in a way that’s possible for humans.

So what are the lessons here? So first off, if something is long, don’t be scared. So if it’s long, first look for the organization. So before diving and trying to find an exact page you need, look at the table of contents. Look at any quick references. Maybe look at the index but that’s not quite the organization you need. But get the big picture of it before you go and try to narrow down to the place to start because the first page you find or a search engine brings you to might not be the actual right starting place. Know that not everything is needed all the time. So be able to not be intimidated that way and look more in depth. Ask for guidance. So if you ask me, part of the reason we’ve written this much stuff is not because we expect everyone to go read it all, but that when people come ask us a question, instead of needing to give everyone an individual answer, we can point them at the relevant page here. So the documentation is meant to be used along with us, not only instead of us [though we hope people can give it a try first!]. You can also ask your friends or colleagues, whoever that’s doing the same kind of work, for information on [things] like “where to start?”, what they actually use, and so on. Finally, realize a lot of the stuff in here may be old and may be out of date. So if you’re reading something and it looks a little bit off or not quite correct, well then don’t just go and give up, but use it as a starting point. Maybe come and ask us to update it or ask how it may be old. You might need to add in other stuff you’re finding online with that if you’re able to do that, but it’s better than removing something and next time someone needs to begin they’re starting from the very beginning once again.

Finally there’s something that I’ve been thinking about and have made here. So I made this little diagram of finding information in the past, if you can see. So, in this [hypothesis], in the past we didn’t have all these computer systems and so on. So when you need to find information, you would have to go to this middle layer of, say, your supervisor or the team expert. You have or a librarian or a research engineer, whatever, who can help you go and navigate the things and find whatever it is you’re looking for. And that’s sort of made all the information more manageable. But then these days we have computers, we have the internet, we have search engines, and so on. So since you can find everything yourself, it’s sort of expected that you [do] try to find everything yourself. And if you ask me this is not necessarily good. I mean, it’s good that we have access to information, but since there’s this idea “oh, we don’t need the team experts, we don’t need people like that anymore, just let people find it all themselves” leads to information overload and there’s just too much for people to be able to navigate themselves and actually slow things down. So make sure you keep this in mind and use your support networks. Don’t try to go alone.

If you’re writing information, if you’re writing documentation for your own equipment, don’t go and just do it without paying attention to the organization. I’d say make a good table of contents, and try to have the table of contents be all on one page so it’s easy to navigate and see everything in it, instead of having to click through different things. It’s like if you make a really deep hierarchy, where to know what’s at the lower levels, you have to click down through several upper levels to see [what is] there. That makes it a lot harder to get the overview and know if you’re going to the right places or not. I like to try to have one page where it’s possible to do a [search] with Control-F within a single page and get a summary of where some keywords are, and from there it should link you to the details. So first have a quick summary there which is useful for quick reference (like in our cluster documentation) with links to the details. So that way if someone genuinely knows what’s going on, they can just use the quick reference and copy what they need. If they’re new they can get the overview and go to the more detailed pages.

So with that being said I need to actually clean up this stuff now. So thank you for watching and see you perhaps in the next low quality HPC kitchen video. Bye.