Today at Open Compute Project (OCP) 2018 in San Jose, California, Toshiba Memory America Inc. announced the release of its new NVM Express over Fabrics (NVMe-oF) shared accelerated storage software, KumoScale. This new software will help companies to maximize their use of NVMe flash drives, and the performance benefits of directed-attached NVMe, over the data center network. The company is also collaborating with Portworx to bring the industry’s first Kubernetes integration with NVMe-oF target. We have also begun testing a KumoScale solution with the Newisys division of Sanmina and the results have been so stunning, we didn't want to wait for the review to complete to share some of what we're seeing.
Today at Open Compute Project (OCP) 2018 in San Jose, California, Toshiba Memory America Inc. announced the release of its new NVM Express over Fabrics (NVMe-oF) shared accelerated storage software, KumoScale. This new software will help companies to maximize their use of NVMe flash drives, and the performance benefits of directed-attached NVMe, over the data center network. The company is also collaborating with Portworx to bring the industry’s first Kubernetes integration with NVMe-oF target. We have also begun testing a KumoScale solution with the Newisys division of Sanmina and the results have been so stunning, we didn't want to wait for the review to complete to share some of what we're seeing.
The cloud was built on direct-attached SSDs for performance, but many are finding that this lacks flexibility. Last year, Toshiba introduced its NVMe-oF technology and was the first to get it certified at UNH-IOL. This software aims to address the above concerns by managing storage functionality, enabling the creation of networked storage nodes that can be immediately deployed at scale – delivering improved utilization of powerful NVMe SSDs by allowing them to be shared. KumoScale can leverage dynamic orchestration for more efficient compute nodes.
One other interesting benefit of the software is its ability to enable high-performance storage for container orchestration frameworks, such as Kubernetes. Storage companies for Kubernetes, such as Portworx, can leverage KumoScale to make containerized big data, fast data and machine learning workloads easier to manage and higher performing. According to the two companies, running Portworx volumes on top of KumoScale’s target namespaces enables the raw performance of NVMe to be efficiently and flexibly configured under the leading orchestration framework.
TMA and Portworx will jointly demonstrate KumoScale at OCP booth #A21 at the San Jose Convention Center from March 20-21.
Review Preview
We are currently in process of reviewing the Newisys NSS-1160G-2N dual node server with KumoScale software installed. The Newisys server is presenting storage via two 100G Mellanox cards with 8x Toshiba NVMe SSDs on one node, the second node is used for management purposes. Load generation is coming from a single Dell PowerEdge R740xd that is directly conencted to the box over dual 100G Mellanox ConnectX-5 NICs, running our standard VDbench profiles. For this preview, we have selected 4K and 64K workloads to share.
Measuring peak bandwidth, we saw a peak of over 13GB/s, utilizing 8 NVMe devices shared from the storage node. Latency was minimal as you can see, topping under 450us.
Sequential write performance was fantastic, measuring a peak of just under 9GB/s, with latency never going above 450us.
Random small-block performance in our 4K test was incredible. We saw 2M IOPS before the latency pushed over 150us, and a peak value of 2.98M IOPS at 260us.
Random small-block write performance was also excellent, with low-load latency measuring under 50us up to around 1M IOPS, then slowly ramping to 225us at a peak of 1.93M IOPS.
As noted, this is simply a preview of the initial data as we work through a more complete review. The latency numbers are very impressive however, illustrating the potential for NVMe-oF when executed well.
Sign up for the StorageReview newsletter
