GPU Died. Training Didn't: Self-Healing Training at Scale — Crusoe

AI Engineer · 16 min · 141 sentences · from YouTube's caption track

Each timecode opens YouTube at the start of that sentence. Line anchors (#s42) are the cue ids in the WebVTT, and every line carries its start and end seconds. All transcripts has every talk, and the whole corpus as one file.

00:01[music]
00:12All right.
00:13Hello everyone.
00:13Uh, my name is Connor.
00:14I'm a developer advocate at Crusoe and I'm joined by my colleagues Nikil and Young.
00:19And we're going to be talking today about the infrastructure we've built at Crusoe and how it handles critical hardware errors.
00:26And so our customers when they're running extremely largecale training workloads across thousands of GPUs, GPU failures are inevitable.
00:35And so at scale, manual remediation is completely unsustainable.
00:40And so what we'll cover today is how by using both Slurm and Kubernetes, we can provide a robust platform with automated remediation that maintains ease of use for both machine learning engineers running jobs
00:53and the platform teams managing the underlying clusters.
00:58So to provide a little bit more context on what Cruso cloud is for those of you who haven't heard of us um we provide infrastructure as a service consisting of compute storage networking and the latest gen GPUs from Nvidia and AMD
01:13but the focus of this talk will be on the tooling that we've built on top of that to orchestrate these large fleets of GPUs including our managed Kubernetes service called CMK
01:25and our manage slurm service which is built on top of CMK and built off of Slinky which is SCADMD's official open source project for running slurm on top of Kubernetes
01:37and this allows us to get the high performance job scheduling of slurm with the infrastructure resilience features and observability of Kubernetes including one of the systems we've developed called autoclusters
01:50which is our system for automatically detecting and replacing failed GPU nodes when a GPU failure has been detected.
01:58And so to explain how and why we landed on this architecture, it's helpful to understand where traditional slurm both excels and falls short, which Young will get into now.
02:13Thanks, Connor.
02:14Um, traditionally, I mean, looking at it from the tra traditional perspective,
02:21sorry, can can you hear me now?
02:22Okay, cool.
02:23I I'll just do it like this.
02:24But basically, traditionally, slurm was built slurm was built by the researchers for the researchers, universities, labs over 20 years ago.
02:35Uh built for high performance computing uh which as we've observed with our customers and others that are using slurm today translates very well to modern AI specifically around training workloads.
02:47this idea that when you're running a multi-node training job, you need this idea of tight collective communication across the ranks that that you're using.
02:58Uh including, you know, specific network built for training workloads today.
03:04Um concepts like gang scheduling, topology awareness, uh prologue and epilog for cluster validations.
03:12These are all the features that are readily available with slurm today that helps researchers run their jobs.
03:20Uh essentially running their jobs with tools that are already familiar to them.
03:23Scripts, especi cluster today.
03:31Now, where we feel like this is a bit of a shortcoming or short falls a little bit short around the modern training is the things that you see on the screen right now.
03:41Um AI training workloads today are very dynamic and it doesn't just include training.
03:47Uh there's a lot of other aspects that your AI labs and other companies are running as part of your your AI workloads including training post-training eval inferences some of which may or may not even land on your slurm cluster today.
04:02But because of of what we've observed as as a GPU capacity constraints and other things kind of limits you to how you need to be able to be more dynamic around the resources that you have.
04:13Uh slurm traditionally is pretty static and it's partitioning and other other uh functionalities.
04:20The other thing I want to mention is this idea of health checks and maintaining your nodes.
04:26Uh you know you can make sure your nodes are healthy or not healthy.
04:30GPUs do fail, how do you manage that?
04:33Uh, SlurpM obviously has features around it to be able to, you know, reue the jobs, uh, detect bad nodes, uh, using things like prologue and other open source tools that are now available
04:45to be able to do this.
04:46It just adds an operational burden, uh, that we've seen from customers today.
04:52you know, this idea that you have to detect if nodes are bad, you have to drain the nodes when they're bad, you have to kind of work around that.
04:59That could either be manual or you need some sort of an automation or scripting to be able to do all of this.
05:05And the last thing is observability.
05:07Um, you know, jobs can fail.
05:09Jobs could also come with degraded performances for any number of reasons.
05:14I mean, GPU failure is just one of them.
05:16there's other things like network flaps and other things within your your switches that may be causing degraded or failed performance.
05:24Uh slurm knows when a job fails.
05:26It doesn't necessarily be able to tell you everything about why it failed.
05:30And that's where the observability comes into key uh key play.
05:34So all of this kind of encompasses the reason why we we built this product that Nikil is going to go over and sort of describe some of the specifics around
05:43manage slur.
05:49Um all right uh thanks young.
05:54So yeah, I'm going to be talking about our Cruso manage slurm product and uh like like we've mentioned before um there's a lot of benefits uh that slurm can bring to the uh training teams
06:08and uh researchers that are building the next best models but there are some issues with you know setting up slurm managing it keeping it reliable uh and on the other hand you have kubernetes
06:21which is the OS of the cloud It's widespread use and familiarity.
06:26It's a very mature ecosystem with lots of tools, extensions, and plugins for whatever you need to add, whether it be observability, networking, security.
06:37It helps uh maintain services with there, you know, built-in safe healing, uh load balancing and autoscaling that keeps everything running smoothly.
06:45And so we often see that uh customers will uh come with us starting with say inference running on Kubernetes and then they want to expand to training uh or vice versa.
06:57And so maybe a Kubernetes familiar team struggles with deploying and managing a slurm cluster.
07:04Um or uh the incoming slurm native training team is unfamiliar with the existing Kubernetes frameworks uh and adds you know additional friction that just slows down your team and with you know how fast everyone's trying to get to
07:19uh the next best model and next best product.
07:23um teams get set up with the tools they're familiar with and you end up with two separate infrastructure stacks which has its own host of extra operational burdens uh and extra costs.
07:34So we decided to build out our Cruso managed slurm uh on top of Kubernetes.
07:40So we have our Cruso slurm operator uh which coordinates with our Cruso managed Kubernetes platform and the broader Cruso cloud.
07:48So including storage, networking etc. And uh cso manages uh slurm users, slurm partitions, configurations and storage.
07:58So you don't have to do it peacemeal.
07:59You just you know one one command you create your slurm cluster and there it is existing in your kubernetes cluster.
08:06So the benefit is for the slurm users um it's just another slurm cluster.
08:11They just need uh an IP address they can SSH with their uh user and now to them it's a slurm cluster with all the bells and whistles that they need.
08:19They don't even need to know that Kubernetes is running underneath.
08:22For your platform team and your uh infrastructure team, uh the benefit is slurm is now just another service that plugs into your existing infrastructure.
08:30You don't need to set up an entirely new infrastructure stack, new observability tools, new you know on call teams, new runbooks, nothing like that.
08:39It's all built into the same thing.
08:40It runs alongside all your other services and um resources in slurps.
08:47any jobs make use of the uh built-in topology aware scheduling um and the platform team just sees it as another service um and now since all your GPUs and hardware are in one cluster they can move around seamlessly
09:01uh it enables uh more interesting uh tooling and dynamic availability like when you need to burst your uh inference service when you're reaching uh high uh users high number of users
09:15you can take away some some of your or even reallocate some of your hardware from your training cluster to your inference service or vice versa.
09:22You know, when your user uh load goes down, you have all these GPUs sitting idle.
09:28That's the perfect time to up your training job or do a a larger hero run across more GPUs.
09:34Um this is something that's only possible when you have everything in a single uh infrastructure stack like uh cru managed slurm.
09:42Um an additional benefit is uh now crucial managetorm can make use of our autoclusters product which is our auto remediation for additional hardware errors.
09:52So Connor is uh going to walk through that and show a demo of uh of how that works.
10:03So this is an overview of what exactly happens when autoclusters detects a critical hardware error.
10:10in this case an XID79 error where one of the GPUs is completely unusable.
10:15And so the first thing that happens is that the user is notified although there's no action required from them.
10:21Then the slurm operator sets that that specific node to down and cancels any running jobs.
10:28There's then a uh the process re receives a sig term signal before being terminated and you have up to two minutes to handle that sig term to either save your checkpoints or flush your logs
10:39before that node is eventually replaced.
10:42The cancel job is automatically reued and then autoclusters start.
10:48So first it verifies that no pods on the node have a label that disables automatic replacement.
10:56Assuming that node can be replaced, the Kubernetes node is then cordoned and drained.
11:02And then the unhealthy node is removed from the node pool and replaced with a new healthy node from spare capacity.
11:10And then a record of the alert and the remediation is logged for the user's reference.
11:15Uh but there's still no action required from them.
11:17And then once that healthy node comes back online, then the slurm job starts.
11:22Your application code runs to load your model and load your checkpoint and resume training.
11:30And so here's what that looks like um in a demo.
11:36And so here on the left, I'm just using sbatch to launch a PyTorch training script.
11:42On the right is the Cruso cloud console where I have a GPU node pool of two A100 nodes.
11:49And I'm going to launch this as normal.
11:53Um, and then what I'm going to do is run a script in to tell our backend internal monitoring system that there's been an XID79 error on one of the GPUs.
12:04And this will trigger autoclusters.
12:06And so we can see immediately the GPU utilization drops.
12:11And so we know that that that GPU has gone offline and autocluster's node replacement starts immediately.
12:18And the full process from detecting a critical hardware error to getting a healthy node back into the node pool takes roughly five minutes with autoclusters.
12:29The rest of the time is mainly application level code.
12:32And so in this example, it's a little under 15 minutes um from end to end.
12:38And here I'm just showing some of the alerts that uh the user receives when there's a critical hardware error like this.
12:45So as you'll see um there's no action required from the user.
12:50The slurm job will restart.
12:52It'll load the checkpoint and continue training where it left off.
12:56And we can see the GPU utilization returns back to normal.
12:59And the total downtime is less than 15 minutes.
13:03And this is um much better than having an engineer have to log on in the middle of the night and spend hours debugging something.
13:11All of this is handled automatically on our platform.
13:18Now, historically, creating this type of environment has been challenging to merge both Slurm and Kubernetes.
13:25And so we're really excited to announce what we call one-click slurm where everything that we've talked about in this presentation, this entire environment can be provisioned with with just a single command.
13:37And so this one command includes provisioning the underlying Kubernetes cluster, the slurm controller, login nodes and storage.
13:45And then adding a GPU node pool is only one command beyond that.
13:51And so this means that engineers can simply SSH in and start running jobs immediately and autoclusters is enabled by default.
14:02So wrapping up what we covered um starting with the limitations of traditional slurm it gives you the high performance job scheduling but not the automated the automated healing that modern AI infrastructure requires.
14:14And so by building slurm on top of Kubernetes, we can enable these advanced advanced features without having either team change uh their workflows.
14:25Then we covered the full node remediation flow from an error being detected to the slurm job resuming and the checkpoint and the training resuming.
14:34And so the key takeaways here first is that traditionally slurm and the infrastructure that it runs on are two separate systems.
14:43But together with Kubernetes, both systems stay in sync and so a failed GPU node can get cord in automatically and that signal will propagate directly to the slurm operator.
14:58Also, like I mentioned earlier, neither the platform teams nor the machine learning teams have to change their workloads.
15:06Researchers can still use the same commands to launch slurm.
15:10And although Kubernetes is running underneath, researchers never have to SSH into a pod or worry about containers, but the Kubernetes team still get access to the same level of
15:21telemetry and observability that they're used to.
15:27And then lastly, like Nicole mentioned, the GPU nodes are Kubernetes nodes first.
15:32And that means that slurm is just a workload running on top.
15:37And when a slurm job finishes, those GPU nodes don't have to sit idle waiting for another job.
15:44You can schedule inference pods on those immediately through Kubernetes because from the platform's perspective, they're still just nodes in a pool.
15:55And so the core idea here is that failures are inevitable.
16:00Um and so in today's AI landscape, the architecture of infrastructure should be designed in a way so that when there is a critical error, all the right actions are handled autonomously
16:12and engineers can focus on building their application and not worrying about their infrastructure.
16:20So hopefully this was an informative talk.
16:22Um, if you have any questions, feel free to um, connect with us at booth ug12 or on social at crusadev.
16:30And uh, thank you for listening and feel free to ask any questions.
16:34Thanks.