Melpomene - Fotolia
Excelero is extending the deployment options for its NVMesh software-defined storage to the public cloud and, with the initial option, targeting some of the most demanding workloads running on Microsoft Azure.
The startup's NVMesh is designed to pool, share and protect NVMe-based flash storage and boost the IOPS and throughput of I/O-intensive applications such as databases, data analytics, AI and high-performance computing (HPC).
Excelero, which is based in Tel Aviv, Israel, claimed benchmark tests on InfiniBand-enabled Azure HBv3 virtual machines (VMs) showed that NVMesh could enable up to 25 times more IOPS and 10 times more bandwidth, while lowering latency by 80%. Azure's H-series VMs are the fastest and most powerful the cloud provider offers, and Excelero used the optional high-throughput remote direct memory access (RDMA) interfaces for the tests.
Support for other major clouds
Excelero CEO and co-founder Yaniv Romem said his company added support for NVMesh use in the Azure cloud in response to requests over the past 18 months from customers who want to move I/O-intensive workloads to the public cloud. Romem said Excelero plans to expand support to other major public clouds later this year.
Marc Staimer, founder and president of Dragon Slayer Consulting, said NVMesh could help customers better use and share the costly NVMe flash storage they buy in public clouds, as well as eliminate any unused capacity they may have overprovisioned to meet mission-critical application needs.
"The Excelero advantage is that they can aggregate the drives from multiple nodes, making all drives in the nodes appear as shared storage with local performance," Staimer said. "This eliminates expensive orphaned storage and enables better utilization at higher performance."
High-performance storage tier
NVMesh must run on all the VMs in a server cluster to pool their NVMe SSDs into a high-performance storage tier. When the flash storage and compute are converged on the same node, Excelero uses its patented remote direct drive access (RDDA) technology to bypass the CPU and speed access to NVMe flash drives. Excelero's internal RDDA protocol originally supported only RDMA to connect the drives in the cluster, but it now also works over TCP/IP, Romem said.
Servers that don't have compute and storage on the same node can access Excelero's NVMe flash pool via TCP- or InfiniBand-based NVMe-oF. In those disaggregated scenarios, customers run their applications on dedicated compute nodes and use separate storage servers equipped with NVMe SSDs and NVMesh software.
Romem said Azure's fastest H-series instances are suitable for converged or disaggregated use, but the converged option is more likely, given the abundance of cores available. He said customers can choose between RDMA- or TCP/IP-based access, but Excelero's RDDA would deliver the highest performance across RDMA for the InfiniBand-enabled H-series VMs.
Customers have a variety of other VM options, including the Azure N-series for GPU-based workloads, such as graphic rendering and video editing, and the L-series for storage-optimized use cases, such as SQL and NoSQL databases and data warehousing. Romem said the N-series has no local SSDs and could either run the NVMesh client or use NVMe over TCP or InfiniBand to consume storage from pooled drives on the H- and L-series.
Azure's L-series can run in converged or disaggregated mode, but Romem said separate compute and storage nodes are more likely with less CPU power available. Storage access would be through RDDA-TCP or NVMe over TCP. Romem added that connection options such as iSCSI and NFS are possible for customers who don't have operating systems with NVM-oF support and don't want to install NVMesh clients.
Customers who run NVMesh both on premises and in the cloud could find advantages for HPC workloads. Mark Nossokoff, a senior analyst at Hyperion Research, said NVMesh is designed to integrate with a user's NVMesh-based on-premises infrastructure with no code changes.
"HPC users are increasingly moving more and more of their workloads to the cloud. One element driving this trend is the ability to burst to the cloud to minimize queue times when on-prem resources aren't immediately available," Nossokoff said. "This could be due to the resources being utilized by other workloads, or the on-prem system has reached full capacity. In order to burst to the cloud, users typically must modify their application codes to support different types of cloud resources than what's implemented on prem. The ability to burst to or pool the same resources in the cloud as is on prem to alleviate application code changes and prevent overprovisioning would be appealing to a large class of users."
NVMesh on Azure pricing
Based on current price sheets, NVMesh on Azure starts at $8.99 per hour using 11.4 TB SSDs running the Excelero software on Lsv2 volumes across TCP/IP. Pricing for NVMesh on Azure's HBv3 volumes across RDMA starts at $12 per hour.
Cloud customers can buy NVMesh through the Microsoft Azure Marketplace, work with Excelero to provision NVMesh on Azure, or install NVMesh using IBM Red Hat OpenShift for use in a Kubernetes environment.
"Excelero is not leading the industry here; although, they may be able to make a case for why their software provides a lower cost, high-performance option than a customer licensing NetApp Cloud Volumes Service or Pure Storage Cloud Block Storage and running it in the cloud," said Eric Burgener, a research vice president at IDC.
Burgener said Excelero's pending addition of support for the leading cloud provider – Amazon -- and Google would be good for Excelero customers. He noted that Amazon bought E8 Storage, another NVMe-based flash storage startup, in 2019.
"The flexibility of NVMesh to accommodate heterogeneous hardware lets a variety of different public cloud providers, who may use different x86 servers in the web infrastructure, select Excelero and use it," Burgener said. "For customers who can now deploy Excelero in the public cloud, they may be able to move workloads that otherwise could not have gone to the cloud, while public cloud providers might be able to advertise a more 'enterprise-like' storage service than they've had in the past."
Carol Sliwa is a TechTarget senior writer covering storage arrays and drives, flash and memory technologies, and enterprise architecture.