KumoScale Software Insights: Thin Provisioning Increases Cloud Storage Efficiency

Wondering why your cloud storage applications are plagued by inefficiencies? First things first: let’s look at existing cloud storage architectures that use racks of compute nodes with direct-attached storage – and the problems they cause. A typical cloud-native data center application might use 10s or 100s of compute nodes when lightly loaded but must scale up to 1000s of compute nodes during demand peaks. Work is dynamically assigned to available compute nodes. Everything is measured on a per-node basis including floor space, power, CAPEX, and storage capacity. For storage capacity, how do you know what size SSD you need in each compute node? For constantly changing workloads, whatever size SSD you choose will likely be wrong. It is difficult to predict storage capacity needs across a range of applications. Modern containerized applications make this problem even worse. This problem is very real and consistent with feedback I get from many customers who offer cloud services, software as a service, or infrastructure as a service.  They complain about very low utilization of their available flash capacity – directly attached, inside the compute server – which was sized to meet peak demand requirements. Most of the time, these customers are reporting that this direct attached storage utilization is as low as 5 to 10%. 

That’s a ton of wasted storage.

So, how do you solve this problem? Shared storage accessible over the data center network helps address this issue. The primary value proposition of storage disaggregation from compute servers is to optimize utilization of both the compute and storage resources. As discussed in prior posts, KumoScale^TM software combined with orchestration software like Kubernetes® enables the most efficient virtualized storage solution for both bare metal and containerized applications. KumoScale software virtualizes and provisions a giant pool of network-attached, high-performance flash storage on demand using NVMe-over-Fabric^TM networking technology. By partitioning the storage pool into thousands of virtual SSDs, each compute nodes gets exactly the needed capacity per application QoS requirements – and the storage appears and performs as if locally attached.

Related to this core virtualization and provisioning functionality, KumoScale software also supports a variety of robust data services to manage flash storage resources at data center scale – including volume replication, snapshots, and cloning. The latest data service addition to the KumoScale software suite is thin provisioning. Thin provisioning enables maximum flash utilization and reduced cost for disaggregated cloud storage, while retaining low latency performance.

Traditional storage systems allocate capacity beyond current needs in anticipation of future growth. The practice of ‘thick provisioning’ often reserves large amounts of flash capacity that may never be used. With KumoScale thin provisioning, capacity can be easily allocated to servers on a just-enough and just-in-time basis essentially eliminating the problem of wasted capacity. Thin provisioning flags can be set when provisioning the storage volume for KumoScale software volumes, and allocated storage volumes can be automatically expanded as needed. By doing this, you can save between 60-90% of your flash cost, which at scale can add up to significant cost reductions (based on reports published by hyperscale cloud providers like Google). KumoScale software’s thin provisioning feature further maximizes storage efficiency and reduces flash cost in a disaggregated storage architecture. 

The bottom line is you can get more work done with the same resources. Or, if budget reductions are needed, you can get the same work done at a much lower cost.

 For more information, please visit kumoscale.kioxia.com.

Notes:

NVM Express, NVMe and NVMe-oF are trademarks of NVM Express, Inc.

KUBERNETES® is a registered trademark of the Linux Foundation in the United States and other countries, and is used pursuant to a license from the Linux Foundation.

All company names, product names and service names may be trademarks of their respective companies.