text stringlengths 0 2.08k |
|---|
ZFS I think is one of the major ones now, and then Slurm is used all over the place, and that was invented at Livermore. That's a resource manager... It's basically like a batch system for submitting jobs to clusters and managing the nodes and the memory allocating time to users. |
**Mikeal Rogers:** These are pretty high-end use cases... Who are the other users of these open source projects, other than Lawrence Livermore? |
**Todd Gamblin:** Other labs deploy Slurm, and actually a lot of university computing centers will run that. Our Linux clusters all run Slurm; even our IBM Blue Gene machine - which is like a million and a half cores - runs Slurm, and other national labs also run Slurm on their systems. |
**Nadia Eghbal:** \[11:09\] The name makes me laugh every time. |
**Todd Gamblin:** Yeah, I think it's actually from Futurama... So yeah, we have lots of interestingly named tools. And then ZFS I think is used in the industry. I'm not a file systems expert, but I know that a lot of companies have started using ZFS on Linux; it's fairly widely used all over the place. |
We ported ZFS because our interest is in -- so we also have developers who work on a parallel file system called Lustre, and that's what we run on our clusters. Lustre is a parallel file system where there's a local file system that it's based on. Lustre runs on top of ZFS in our current configuration, and we're pretty... |
**Nadia Eghbal:** Was Spack the first project that you had open sourced at Lawrence Livermore yourself? |
**Todd Gamblin:** \[12:03\] It wasn't, actually. In addition to Spack, I've had a bunch of different research projects. For my PhD I worked on a scalable clustering algorithm. That's open source, it's called Muster and it's on GitHub, too. That was sort of for finding nodes in a parallel application that very similar p... |
I don't think that really caught on... It was sort of a research project; it wasn't generally useful, like a package manager is. We also had a project here called -- well, originally I wanted to call it ClownCar, but that was deemed not serious enough, so we renamed it to CRAM, which stands for ClownCar Renamed to Appe... |
**Nadia Eghbal:** Oh, government... |
**Todd Gamblin:** So that dealt with a real problem that we had here. When we deployed the Sequoia machine, which is the big IBM Blue Gene/Q system, originally people anticipated running maybe like 200, maybe 400 jobs at a time on that thing, and they were thinking that each job would be maybe 10,000 cores, or somethin... |
Now the mission has sort of changed, and one of the things that we're very interested in is uncertainty quantification, so trying to figure out what inputs is the simulation sensitive to? We decided that we wanted to run lots and lots of small jobs on the machine; we had people who really wanted to run a million and a ... |
CRAM basically takes one job and splits it into lots of them, and manages that sort of on the cluster, as opposed to on the front-end. So it's kind of a stopgap until we get a more scalable resource manager. |
Those are just some examples, but I would say in the research world there's lots of open source software. That's kind of the default for people who are doing research, publishing papers about it... We have a compiler project here called ROSE - it's a source-to-source compiler; that's been open source for a very long ti... |
**Mikeal Rogers:** Do you have a sense of how long it's been the default? Because it hasn't always been the default to do everything open source. Do you know when that shift happened? Or just like as long as you've been in it, you've been able to do everything open source? |
**Todd Gamblin:** As long as I've been in it, yes. I started here at the lab in 2008, and I know that it goes back further than that. Our early efforts with TOSS, which our Linux distribution (based on Red Hat) - that was in '99 or the early 2000s. Actually, there's a policy document from the DOE from 2004, and it basi... |
An open source nuclear weapon simulation sounds like a really bad idea, but some of the other software that we develop to make our systems run - that's pure computer science; we can put it out there, and other people can use it and benefit from it... And they do. |
So yeah, there's a DOE document that says we should make this software open source from like 2004, and it's interesting because -- you know, that's something we're struggling with now... We do have an IP organization as part of the Laboratory. Livermore has had some successful commercial software come out of some of ou... |
\[16:20\] But other projects, like Spack or a lot of this infrastructure stuff - I don't think we're gonna be able to sell that, and I don't necessarily think that we should, because we're in sort of a niche here. There aren't that many sites that do large-scale high-performance computing, and I think we have the same ... |
**Nadia Eghbal:** So what's it like when you think you have something that you might wanna open source? What does that process look like for your Lab? Do you have to talk to the IP Office? |
**Todd Gamblin:** Well, yeah... So the IP Office wants to make sure that the thing that you're putting out there as open source is not something that we could potentially get royalties off of. That's kind of interesting, because the government in general, technically -- I don't think we're actually allowed to have inte... |
But yeah, so we have to go through the IP organization when we put things out there. For most of my stuff, getting the approval isn't actually that hard; it's the actual process that's kind of tedious. Livermore's software release process involves burning two CDs and carrying them to three buildings throughout the Labo... |
The thing that I would really like to do here is start thinking about open source a little more carefully in terms of like "What is our open source strategy with this thing?" Look at open source as a potential software sustainability strategy for DOE projects. That's kind of like what we're doing in Spack; we've active... |
Depending on what it is, maybe that's a good strategy, maybe it's not. There's some things that we wanna develop internally that we wanna keep a well-staffed team on, and there's other things that we can probably share effort with other laboratories on and that we would wanna build like a larger community maybe, even o... |
I was on a working group here that was trying to come up with -- I guess it started out as like a software engineering working group, and I think by the end of it we had changed the name of the working group to Software Sustainability. I think everyone thought we were gonna come back and say "Okay, everyone should do w... |
**Adam Stacoviak:** After the break, Nadia and Mikeal talk with Todd about the back-story of Spack and where it's going. We talked about Spack's growing community, how they're finding contributors, especially in a government context. We also talked through the details of Spack becoming a NumFOCUS affiliated project and... |
**Break:** \[20:15\] |
**Nadia Eghbal:** Todd, you've mentioned Spack a couple of times as an open source project that you've created while at Lawrence Livermore... Could you explain a little bit about what Spack is for people that are not high performance computing people, and what inspired you to create it. |
**Todd Gamblin:**\[21:47\] I have an easier time explaining Spack to people who are not high-performance computing people than to people who are high-performance computing people, in the sense that like I think people who are not high-performance computing people are familiar with package managers... Spack is a package... |
I think the reason that I built my own package manager is because the way that we distribute software in HPC is kind of different from what you would expect from most software products. People actually build their code from source on the supercomputers, typically because they wanna optimize it for hardware, but also be... |
If you look at the climate community, they have just a whole bunch of Fortran models that are tied together in different ways. Our codes tend to be like C++, some C and maybe some Fortran down in the numerical libraries, but primarily they're C and C++. We have Python drivers for some of those things; I think recently ... |
But yeah, so one simulation could be 47, 70, maybe 100 libraries in C, C++, Fortran and Python. So building that and linking it all together and getting fast native libraries there is kind of hard. |
And the other thing that we care about is exotic compilers. So we don't just build with GCC, we build with GCC, Clang, Intel compiler, Portland Group compiler, and then maybe the Cray compiler on the Cray machines... Lots of different ways to optimize your code. And there really wasn't a good package manager for experi... |
\[24:05\] Another problem with the software ecosystem here is that people will distribute their code, and maybe it relies on a very specific version of some research library, because maybe that research library is published like by a university, maybe they don't have the greatest release cycle set up, or maybe they don... |
Essentially, Spack is a tool to build software the way that I was building it, where I would have lots and lots of different versions of things, and essentially any dependency graph that you make is a new version, so we assign a hash to that. And that's a lot like -- if people are familiar with Nix or Guix or some of t... |
The idea is that you would be able to make a package that could build across all those different platforms and that people would be able to install it and say "spec install foo" and have foo, which is not generally the experience on supercomputers. |
**Nadia Eghbal:** Interesting just hearing some of the things that you have to be particularly concerned about, and it feels almost like you're doing open source but for a smaller but more engaged community of users, beyond your Lab. It sounds like a lot of it is for other labs, or that type of audience, more than a ge... |
**Todd Gamblin:** Yeah, I mean... I think for a lot of the simulation codes, the usage model is very different from what you see in web development. Depending on the team -- some of the teams are their own users, so it's a research project, they're developing this thing and they're testing out different versions of it ... |
Other teams do publish software that can be used by lots of people and they'll actually bother to package it themselves, but I think the build process on these machines is so complicated that your typical computational scientist doesn't wanna get into all those details; they wanna focus on the science. |
We actually talked to different teams in the community and try to figure out "What's your deploy process?" Some of them had good practices, and others said "Well, we have one user on every machine, because we build the code once, and there's this guy who runs it for everyone." Then everyone talks to him, they send him ... |
Now we're starting to see more sharing in the community, and also this sort of push for exascale computing has caused people to really think about how portable their code is... So things like this have started to matter to them. |
We have people who are really concerned with performance portability, which is another good reason to package your software and to use libraries that work across different architectures... For a long time - since like the '90s - if you could run on a Linux cluster you were probably okay and the processor architecture o... |
\[28:14\] A GPU is very different from a Xeon Phi in terms of how you would parallelize code for it, and it's very different from your multicore processor. So I think people have started to realize that they have to rely on libraries for that. They really wanna separate that concern out and give it to someone else who ... |
**Nadia Eghbal:** So how does that play into thinking about getting contributors for an open source project like Spack? Because I imagine -- I noticed that Spack has a lot of contributors, and there are only so many people that are using Spack in the first place, or your total audience is smaller... Do you feel like yo... |
**Todd Gamblin:** I think that one really intriguing place where we could get more contributors would be like industry... I think definitely we could grow the audience for it right now. I think there's a lot of people who still just build things by hand. |
I think when you look for contributors you only have to think about what's the structure of the community. For HPC actually I think there's maybe more roles than people are used to in the software community. I think you're used to thinking like developers and users, but actually we have -- so there's users, there's lik... |
Actually, one of the motivations for Spack was that that particular task was getting to be very overwhelming at Livermore computing, where I am. |
In my case - I was a researcher and I wanted to deploy things on the machines for students and post-docs who were working for me who don't necessarily know how to build all these things, but they want to run them. Then we also have a users support team that deals with the application teams, so they deploy software for ... |
The original contributors to Spack were people who were at these HPC centers who were just sick of building things by hand day in and day out, and wanted to deploy things on these machines. Our deployment model is a little different from what you might be used to from the cloud... Most of these machines have like a sha... |
So yeah, we targeted the HPC centers to try to get them to work together initially... But it turned out that we started getting more contributions from actual application developers and people kind of hacking on HPC code on their own, and I think that's actually served to really grow the community, but that wasn't some... |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.