In a world where the only constant is change, promising that your data center can deliver predictable performance is a constant challenge. With different teams driving varying workloads and the added volatility of seasonality (i.e. Black Friday, Cyber Monday), balancing compute and storage capacity in your cloud data center is a constant juggle of worst-case scenarios.
The good news is we’ve found the culprit—direct attached storage (DAS). By tying capacity/node of storage to IOPS/TB of compute, DAS has three inherent drawbacks in the cloud data center that today’s dynamic workloads exacerbate:
- Stranded Capacity and IOPS. With teams that have stateful workloads that are data heavy and typically require dedicated hardware, storage performance is key to their throughput. Examples of this might be an online marketplace or ad serving. In these instances you have to overprovision storage to ensure you have sufficient capacity for peak workloads, despite the fact that it results in unused capacity. Likewise, for stateless workloads like web servers, while your CPU utilization is high, attached storage is minimally used.
- Stranded Compute. Dedicated hardware also means that in some instances, you are stranding compute. For example, applications that store infrequently accessed user profiles require sufficient storage but leaves compute underutilized. Or, in the case of seasonal businesses like retail, for 350+ days a year compute capacity is significantly underutilized.
- Lost Operational Agility & Revenue. Inherently, the need to overprovision limits cloud data center flexibility. In the DAS model, it’s impossible to quickly reallocate resources for new opportunities. While theoretically batch workloads would be ideal for resource “borrowing,” today’s dedicated hardware models make that impossible.
The answer is disaggregation. By independently scaling storage and compute resources, disaggregation in the cloud data center can help eliminate stranded storage and reduce the number of SSDs required. Likewise, companies can purchase fewer servers because compute resources can be dynamically prioritized during demand peaks once applications are no longer restricted by the data locale limitations of DAS. As a result, you can respond to new business opportunities, usage spikes, or reprioritization quickly and easily. And with today’s high bandwidth, low latency networks, mature orchestration frameworks, more cost-effective and capable NVMe™ SSDs, and the new NVMe™ over Fabrics (NVMe-oF™) protocol, your data center troubles can be a thing of the past.
Toshiba’s KumoScale™ changes the game from DAS to disaggregated storage, and create a highly agile NVMe-oF cloud environment.
Kumoscale™ is a trademark of Toshiba Memory Corporation. NVMe™ and NVMe-oF™ are trademarks of NVM Express, Inc.
The views and opinions expressed in this blog are those of the author(s) and do not necessarily reflect those of Toshiba Memory America, Inc.