The Scale in DriveScale

The Scale in DriveScale

2020-06-12T14:53:35-07:00

The DriveScale Composable Infrastructure solution was built with modern application design approaches in mind. Trends have conspired to make the divide-and-conquer approach of scale-out be the prevalent technique to increasing application capacity and performance. We now are faced with the management of applications at a scale not seen in the traditional data center of the past or the “virtualized” data center that we are emerging from. And that was why DriveScale was formed.

I’ve talked before about the trend towards partitioning applications to analyze massive amounts of data to derive business intelligence. This was the first trend driving the use of bare metal industry-standard (aka commodity off the shelf) servers to solve a problem by breaking it into pieces and distributing the work over 100’s or 1,000’s of nodes. Of course, the original classic application example here is Hadoop.

The second more recent trend is the emergence of containers and Kubernetes and the micro-services approach to application design. The difference from the former trend of simple scale-out of data-intensive applications is that it not only does it use the partitioning approach of using many commodity compute nodes to run the application, but encourages microservice design approach where the application itself is broken down into a set of cooperating smaller services. These microservices are containerized using a lightweight standard “virtualization” scheme that cooperates to deliver the overall application service. This cooperating microservices approach, all in containers, is what Kubernetes manages dynamically to provide elasticity, availability, and version control.

What distinguishes both these use cases in production is the scale of managed elements. Unlike management software to configure compute and storage blades in a chassis, instead of tens of managed elements think of 1,000’s of physical compute nodes and each node having 10 or more configured storage devices leading to 100,000 managed elements in a disaggregated data center.

The preceding table shows one long-term production deployment of DriveScale. Managed elements in a disaggregated data center include the compute nodes themselves – or SmartNICs providing secure platform isolation, the individual storage devices (which can be virtualized into right-sized chunks when appropriate), network-to-storage adapters (which in the case of NVME can simply be the fabric/Ethernet adapter to NVME device), and the topology information necessary to provide required degrees of fault tolerance in a redundant, highly connected (Ethernet) network. Layering a logical control plane like Kubernetes on top of the physically disaggregated platform increases the number of managed elements (and element relationships in the configuration) through the lightweight virtualization of compute platforms and adding the pod entity as a basic managed element.

What problems does managing a data center at this scale bring with it? The disaggregated data center achieves cost efficiencies through the use of Commodity Off The Shelf components and standard Ethernet networking increasingly with NVME as its data fabric. Maintaining persistent bindings of thousands of dynamically created instances requires the robust highly-available configuration store that DriveScale provides. DriveScale software is flexible enough to recover configuration operations in progress in the face of transient network and component outages, which are all too common at this scale. Automation of common operations when plumbing the data path between elements to include application-specific RAID, encryption, file system, and transport is provided by DriveScale’s software. Complexity is pushed down into policy-based automation when deploying or refreshing parts of your deployment. Transparent integration with Kubernetes to provide physical layer orchestration and performance insight to Kubernetes’s logical application view through dynamic storage provisioning removes the complexity from deploying large scale containerized applications.

Operating at this scale is a challenge, and what DriveScale does best.

About the Author:

Brian Pawlowski has years of experience in building technologies and leading teams in high-growth environments at global technology companies. He is currently the CTO at DriveScale. Previously, Brian served as Vice President and Chief Architect at Pure Storage, focused on improving the user experience for the all-flash storage platform provider’s rapidly growing customer base. As CTO at storage pioneer NetApp, Brian led the first SAN product effort, founded NetApp labs to collaborate with universities and research centers on new directions in data center architectures.

Leave A Comment