Startup Vexata is tackling a vexing problem: how to eliminate input-output bottlenecks in storage fabrics without requiring customers to rearchitect their IT infrastructures.
“If you look at the world today, all these IoT environments with [machine learning] and AI ecosystems, where the desire is mass volumes of data that need to be accessed with very high throughputs, at super low latencies and massive scale – how do you do this economically? How do you do it simplistically so that operationally those applications become a reality?” says Ashish Gupta, Vexata’s chief marketing officer. Vexata’s storage software platform “is designed to address this emerging application ecosystem problem.”
Vexata’s executive team has amassed an impressive range of storage, compute and networking expertise. CEO and cofounder Zahid Hussain headed the flash products division at EMC, while CTO and cofounder Surya Varanasi held executive R&D roles at EMC and VMware. Combined, Hussain and Varanasi hold more than 70 patents in the fields of networking, storage, computer systems, graphics, and video stream processing.
Vexata emerged from semi-stealth mode in the fall of 2017 after nearly four years of development and announced a $54 million round of funding from Intel Capital, Lightspeed, Mayfield and Redline Capital. At the same time, the San Jose company unveiled its Active Data Fabric platform, an all-solid-state data storage infrastructure that’s based on the company's VX-OS storage operating system.
Vexata’s software and system architecture is aimed at tapping the full performance potential of current solid-state media, such as NVMe (non-volatile memory express) flash drives and 3D XPoint from Intel (Optane SSDs) and Micron.
The focus of all-flash storage systems to date has been to improve performance for legacy workloads compared to hard disk arrays, but that doesn’t begin to maximize the throughput capabilities of an array of flash SSDs, says Gupta (who was most recently with hyperconverged storage startup Springpath, which Cisco acquired). “When you have a stack that’s not ideally designed for this new media, you end up with less-than-efficient utilization of that media, and that’s the challenge that exists in the market today,” Gupta says.
High volume transactional and analytic applications – such as cognitive computing, high frequency trading, risk and fraud analytics – require better performance than what a typical all-flash array can provide. State-of-the-art compute and networking infrastructure can saturate traditional all-flash arrays, leading to greater latency and poor CPU utilization.
Enterprise companies “have done everything they can to reduce latency at the server and the network. But when it comes to the host, the traditional stacks really take many milliseconds of latency when they do any IO. That’s many milliseconds of latency versus, in our case, just microseconds,” Varanasi says. “We make it very seamless to consume your data.”
How Vexata tackles latency
Under the covers, Vexata took a fresh look at solid-state storage architecture. Its core innovation is the VX-OS embedded operating system that powers the family of Vexata products, beginning with the Vexata VX-100 series.
The Vexata VX-100 systems consist of two active-active controllers and can be configured with between four and 16 intelligent and hot-swappable storage blades, each of which contains four off-the-shelf Intel Optane or NVMe flash solid state drives for a maximum of 64 SSDs per 6RU chassis. Administrators can scale out capacity and throughput by adding blades. Enterprise data services include thin provisioning, space-efficient snapshots and clones, and encryption.
Vexata says its system delivers massive throughput with ultra-low latency targeted toward transaction processing and analytics workloads. When fitted with NVMe flash drives, the VX-100F delivers seven million IOPS with latency of around 220 microseconds and bandwidth of 60GB/sec. The VX-100M, fitted with Optane 3D XPoint drives, delivers seven million maximum IOPS, latency of about 40 microseconds and maximum bandwidth of 80GB/sec.
Vexata storage is also available in the form of VX-Stack Appliances, which are preconfigured, application-centric appliances for platforms including Oracle, SQL Server and SAS; and the forthcoming Vexata VX-Cloud, which will allow the VX-OS software to be run on commodity servers and switches in public and private cloud data centers.
CTO Varanasi singled out three developments that distinguish Vexata’s technology and enable its high system performance. First, the team rewrote the storage stack with the Vexata operating system using Intel’s Storage Performance Development Kit (SPDK), “our goal being to reduce the latency of our software to the absolute minimum possible,” he says.
Second, each enterprise storage module, or blade server, leverages an FPGA (field programmable array) that acts an acceleration engine to handle functions such as IO scheduling and system-level garbage collection. In conventionally architected storage systems, these functions can be a cause of performance bottlenecks when performed by the storage controller. “We realized that this is an enormous burden for processors, so we built an offload accelerator,” Varanasi says.
Third, the main controllers and storage blades are linked by Ethernet, which is appealing for its scalability. “So, our I/O scheduling technology, the way we reduce latencies in our stack, and the fact that we use Ethernet, give us both throughput and very low latency that’s critical,” Varanasi says.
Vexata's ease-of-use approach
Another priority was to enable Vexata’s technology to be deployed seamlessly into existing enterprise storage environments, Varanasi says.
“You’re trying to go through all this data, and harvest information, so you can analyze it and determine how you change your business for it, but you’re doing all that with applications that may not be brand new – existing Oracle, or SAS or SQL Server” for example, he says. “You need to just speed them up, flat out, without actually changing the entire ecosystem. That’s where we play. We just plug and play into existing environments.”
Vexata’s ease of deployment stands out to Tim Stammers, a senior analyst at 451 Research. He points out that there are a number of companies pioneering all-NVMe storage – including Apeiron, Attala Systems, E8 Storage, Excelero, Mangstor, Pavilion Data Systems, Pure Storage and X-IO – and a number of tactics used to eliminate the IO bottleneck that occurs when all-NVMe storage is built using conventional architectures.
“The message from startups pioneering all-NVMe flash storage is that entirely new architectures are needed to exploit the performance of the fast-emerging protocol. Vexata is no exception – under the covers, its FPGA-powered system is very unconventional. However, Vexata's device appears significantly easier to deploy than many of its peers, while delivering a claimed performance that outstrips conventional flash systems,” Stammers summarizes in his recent report.
“It's still early for this market sector, and as yet there are no clear trends or winners,” Stammers says. “However, we note the impressive resumes of Vexata's management team, and the fact that it has already named three large customers [online service provider Oath (AOL/Yahoo), contract IT hardware maker Sanmina, and manufacturing and services conglomerate Tata Group].”