Accelerate cloud and AI workloads with the Hammerspace Data Platform

I'm, Dan Rieger, Senior Product Marketing Director with Hammerspace. The, our presentation today will be about accelerating cloud and AI workloads with the Hammerspace data platform. Joining me today is Chad Smith, our field CTO, and, we have, as Alistair mentioned, a history of cloud field days, including some AI field days that we've participated in in the past.

Accordingly, I'll be focusing on, our work with clouds and customers using clouds. If you're interested in more of the basics and the core components of Hammerspace and the story, for the on-prem, environment and work that we do there, there are previous presentations from a few months ago with Molly Presley, Floyd Christopherson, Kurt Kuckein, and others from Hammerspace that you can, see. We have links to those at the end of this presentation.

Uh, but I'll be focusing on cloud and to some extent, hybrid environments as well. So customer challenges. We've been in a cloud world for twenty years at this point.

AWS was introduced back in 2006. Uh, but migrating workloads to cloud is not always simple or straightforward. Sometimes it is, if you're a very atomic workload, if your data is located in a single location, it's well separated from other things.

Sometimes you can just move both the compute and storage for that workload to the cloud, easily and directly. In other cases, your data is being maybe produced at the edge of your network. Uh, maybe it's scattered throughout multiple regions in your environment.

You may have regulatory concerns. There are all sorts of things. Uh, some organizations begin a migration of a workload to cloud without thinking through all, what all the data that workload's going to need and where it is in their environment.

So it's not always smooth and straightforward. And then, of course, legacy storage, whether you were to buy that for your own on-prem data center or the storage that cloud providers are using as the basis of their own storage services, is not always designed for today's workloads in terms of providing the highest levels of performance, in the cloud. So Hammerspace helps address these challenges.

Customers basically want just a few things in this sense. They want to simplify and speed cloud migration. They want to, access data everywhere and anywhere at any time throughout their environment, and they want to maximize performance, which both means the performance of the storage and the performance of the data, but also the workload performance, which can get backed up if it's bottlenecked by storage on the back end.

So this diagram, I could probably spend most of my presentation, my portion of today's presentation, just here on this. In this case, this is an example of, an environment in OCI. So you can see you have a customer's on-premises environment on the left-hand side.

Uh, you have your environment within an Oracle Cloud region, some OCI services over on the right-hand side. And then within the virtual cloud network, you have customer workload, Hammerspace, and then the VMs with block storage. In this case, we've gone ahead and built out.

At a conceptual level, this remains true for any cloud. We're using the icons here that OCI uses. Their own customers and their own company are familiar with those.

We try to do that for each c- cloud provider, but the same kind of general principles apply. So if we take this to the next level of detail, we have a couple... Here I've, I've selected bare metal, GPU servers as the customer workload.

Um, the Hammerspace environment has two parts, metadata services and, servers and data service nodes as well. Then there's some actual storage. In this case, I've got block volumes connected into what OCI calls their flexible virtual machines.

Those are virtual machines that can be, easily reconfigured in a very granular way in terms of their memory and storage capabilities. Um, the, this is kind of the most basic sort of environment that you could put in a cloud. But as your workload grows, your Hammerspace deployment can grow with you.

Here I've added more nodes. Uh, maybe we're starting to get into GPU clusters, doing training instead of inferencing in the cloud. Uh, I've added the Bastion subnet and a few networking details there in the, in the diagram.

Uh, the Hammerspace environment, we recommend that you cluster the metadata server for availability and reliability, and then you can scale out the data services service, servers, and then, of course, the nodes with the storage at the bottom there. But that's not all we do. Y- the picture that I've painted so far is kind of vertical, right?

There's a w- customer workload at the top. There is a Hammerspace kind of in the middle and some storage underneath it. That's not all that we do.

I've now added, both additional workloads, additional Hammerspace, and then the rest of the OCI storage services, because Hammerspace can connect into file storage, file str-- their managed file, storage with Lustre capability, and also, object storage through, an S3 connector. We'll talk a little bit more, both I and Chad will, go into a little bit more later. So we've got a broader array of storage we're connecting at that point, and now I've added in, maybe you have on-prem storage arrays that you'll need to pull some data from into the cloud.

Maybe you have other clouds that you're using and that have their own storage. The thing that I, couple things that I wanna emphasize here is that-Hammerspace is providing a unified global file system namespace across all of this storage. Um, whether it's storage within OCI, we can configure in different ways, or the storage you've got on-prem, we're able to tie that all together into a single view.

And that's part of what really helps address that use case of easing migration of workloads to the cloud. Um, actually, before I, I hit this point here about performance, I'm gonna step back just a little bit to remind you that everything I just described is entirely agentless. The only software that you see from Hammerspace that's running in this environment is within that Hammerspace subnet.

Um, that's where the Hammer- the Hammerspace software runs. Doesn't matter what your workload is, if, if it's, a GPU clusters or just regular servers, doesn't matter what storage you're using within the cloud, doesn't matter what storage arrays you have on-prem, it's all an agentless solution as well. The numbers here represent, roughly speaking, the performance tiers of different parts, of the, your storage infrastructure.

Most directly, the customer's own workload. If, in this case, it's using GPU, servers or other OCI bare metal servers. OCI even has VMs that have direct NVMe in the servers themselves that are hosting those VMs available for customers.

That's gonna be the fastest storage for the, the customer's workload. The next most high-performing storage is the block volumes. Uh, they're attached to the Flex VM servers at the bottom there.

OCI is a particularly, great match for Hammerspace in this regard because their block volumes, they use a single type of block volume that's configurable all the way up to three hundred thousand, IOPS per volume and one point three million IOPS per compute instance. Uh, so that, those block volumes, in fact, even have a performance auto-tuning feature that you, that customers can turn on, set their own bounds and limits for performance, and it, the cloud will automatically change the performance of the volume in response to changing demands. So that's another, great thing.

Then, of course, you've got file storage, with Lustre file storage and object storage. Um, I wanna note that the, connectivity even applies to the storage arrays on-prem. Of course, going out over the internet is gonna take a little bit longer, and then, of course, out to other clouds.

But, Hammerspace's data orchestration capabilities and assimilation capabilities mean that you don't have to migrate data to make this work. That's one of the key, key facts. When we assimilate data, we take the metadata from all the different storage in your environment, and in fact, build and add a d- a little bit of additional metadata, to that.

And then, that is what prevents us from needing to migrate data that you don't necessarily need. So when you stand up a workload, this week you need a certain subset of data from on-prem or another cloud or somewhere else in your cloud environment, that can be moved all the way up to the local NVMe storage in that bare metal compute that we were talking about, before. And then the next week, you need a different set of storage in a different location.

Data can be moved around as needed, but data is not moved unnecessarily. It's an objective-based policy engine that orchestrates data movement. So yeah, to summarize, we, Hammerspace is the data platform for both AI, HPC, and really any workload that you might wanna move to the cloud that deals with a lot of unstructured data throughout a large environment.

The single namespace makes it easily accessible. We have a couple different ways we accelerate performance. I talked about TierZero, the parallel, NFS as part of, NFS four point two.

I don't wanna get into protocols. That's more Chad world. He'll talk about that a little bit.

Uh, and then orchestrating the movement of the data, as well is gonna help you with performance, and, again, that's not just data performance itself, it's also the workload performance because you're able to maximize the use of the compute in the environment. Okay. I've already pretty much talked about the different capabilities that we have.

You can learn more about these on our website. Chad will, delve into several of them as well. Um, but I do wanna pause here for any questions before I step forward.

Um, yeah. Just in case I missed it, I was wondering, does the global namespace, like, does it span across multiple regions? Multiple cloud regions?

Yeah. Yes, absolutely. Thanks.

Yeah. com. Um, so obviously here we're looking primarily at objects, files, PDFs, documents primarily.

What about the OCI capabilities with the Oracle Database itself? So Ha- Hammerspace is, focused on unstructured data. Unstructured.

So we don't, we don't get into Oracle's database world. Gotcha. Uh, that...

If you're running Oracle Database in OCI, great. Uh, it's a great solution, but, that's its own separate world, and you would use their capabilities for data, the database. Okay, great.

Great. All right. Thank you.

Uh, I wanted to mention that we are, targeting, we're kind of, I broke the use cases down into five here, but you can kind of group them into three different areas. Uh, the high-performance scenarios are at the top, and we're obviously suitable for anywhere that you want your storage accelerated because, cloudStorage is great as far as it goes, but you typically don't have things like a parallel file system available to you without engaging in something like managed Lustre, and we offer quite a bit more, than that. You may have also heard about the, the SSD crisis, and that's something that I think we'll see increasingly affect customers in various ways.

On the consumer side, it's already been driving up the cost of laptops and desktop computers. Cloud providers may have a little more insulation there, larger purchasing power, longer term contracts, but I would not be surprised if we started to see some increase in cloud storage costs at some point there. Um, making the use of a company's own internal storage on-prem and maximizing the value of what they're getting out of their cloud storage makes a lot of sense as, as those prices, may end up rising.

So we, we've been thinking about that and, and, what we can do with the product in order to, to, continue to help customers with that. The, bringing data to the compute. It used to be you generate a bunch of data, you'd have data gravity, and you'd want to take your compute to the data.

In the modern world, especially with cloud regions, you, may well want to do the opposite, bring your data to the compute, and that's part of what we do. Uh, there's a sense in which storage has been a little bit behind as a, a solution. It's, years ago, you saw, other companies do things like virtualizing your compute environment so that you could treat your compute as a big resource, divvy it up the way you need it with a bunch of VMs, and distribute those in a standardized way to your environment.

This is bringing a storage approach to the cloud that helps you handle your storage in a more unified way. Um, the last thing I'll tou-touch on briefly and toward the bottom, OCI has something they call Dedicated Region, and it's kind of a product set more than a single product. Um, but essentially, a customer can purchase that from Oracle and deploy over a hundred of the core services fr-from OCI in their own data center, in their own environment, in as few as three racks.

That's the smallest configuration that, that they can do, which is pretty small. And that really helps customers who have concerns around data sovereignty, data locality in terms of their environment, regulatory concerns, et cetera. And we're...

Hammerspace is already approved for deployment and use in, OCI dedicated regions. The, we have customers already using this today. The, Meta is one of our customers.

We've been working with them for about two years. We started on-prem with them and have expanded into cl-- into a cloud solution. They're using thousands of servers and tens of thousands of GPUs, in this particular solution.

They're starting at forty petabytes of data and looking to expand that. Um, the thing that I kinda wanna emphasize here is that, we are working with these customers to create a solution that really maintains a lot of their... what they're already used to.

So we're not just automating and orchestrating data movement. Um, for example, our S3 connector is more than just like a cache sitting at the center of things. It's, it, that's not how it's designed.

Um, it actually is transparent and gives them access to, S3 o-based object storage as a direct part of the file system itself. Chad will talk a little bit more about that. This means that it's very transparent for them to use Hammerspace, and they're not having to change their own IT processes and so forth in order to make use of us and benefit from Hammerspace's, capabilities.

We also have a couple other customers I'll note here. These are household names on the same order as Meta in terms of your, your recognizability. They're sensitive, so we're not able to tell these stories with their brand with them at this time, but we're looking to make that happen.

They're very par-parallel. There's, hundreds of GPUs in the communication space customer with four petabytes of storage plus additional object storage, and the financial services company has been migrating. This is not an AI use case.

And also, I should mention, I've talked a lot about OCI today, but there's actually multiple clouds represented among these customers. Uh, and they, they are benefiting from the same sort of workload automation for the data orchestration as well as the, the application environment as well.