Vexata: A NVMe Fiber Channel array with an attractive price / performance ratio

The result of the work of former EMC and VMware executives, the Vexata VX100 100% Flash array is an NVMe-based fiber channel disk array with low latency, high IOPS performance, and a reasonable price. An explosive cocktail.

In a recent round of US start-ups, LeMagIT was able to meet the team at Vexata, a young company that has developed a 100% Flash SAN storage array optimized to deliver minimal latency.

Founded by former EMC and VMware executives, Vexata has designed a distributed storage architecture that maximizes the performance of NVMe drives without forcing companies to overhaul their infrastructure.

This architecture is currently embodied in the Fiber Channel SAN VX-100 array . The system comes in the form of a 6U chassis that accommodates up to 16 storage blades, each equipped with 4 NVMe SSDs (ie 64 disks per bay). These blades are actually motorized microservers by ARM Cavium processors, whose mission is to extract the maximum performance from NVMe SSDs while delivering optimal latency.

Architecture principle of a Vexata bay

They are interconnected through a redundant Ethernet backplane. On this backplane also come two redundant Xeon-based controllers that run the advanced services of the array (thin provisioning, snapshot, replication, clones). These controllers, powered by the home OS, VX-OS, play the role of a distributed control plan. This control plan is accelerated by a series of FPGAs on the redundant Ethernet switches in the array.

The architecture of the firm allows according to Surya Varanasi, his CTO, to ensure minimal latency. The I / O management stack of the array would have a latency of about 5 micro-seconds, against more than 1250 microseconds for most traditional architectures.

Vexata puts forward an optimized architecture for latency

It should be noted that it will soon be possible to stack several bays via the multiple 40 Gigabit connections present at the back end of each controller, in order to deliver even higher performances.

High performance in all scenarios

By leveraging a distributed software and hardware architecture, Vexata promises high performance even for the most demanding applications. As Surya Varanasi explains, most of the current bays rely on a large cache floor to deliver high performance. While this cache strategy is very effective for applications that essentially perform read operations, it collapses as the writing proportion increases.

The problem, according to Vexata, is that most modern analytics workloads, as well as emerging applications of machine learning or deep learning, have I / O profiles that are radically different from historical applications, with a read / write ratio close to 50/50. . This evolution literally collapses the performance of historical storage architectures.

As noted by Varanasi, a Dell EMC VMax array often has no issues in delivering the expected performance for critical ERP applications with read / write ratios in the range of 90/10. But if the ratio falls to 50/50, then we are far from the 4 to 6 million IOPS promised for a high-end bay, but rather close to 250,000 to 400,000 IOPS. More seriously, data access latency explodes.

It is to cope with this evolution of applications that Vexata has designed its architecture, explains the CTO. According to him, a Vexata array maintains its performance, regardless of the type of I / O profile of the workloads it faces. With an Oracle OLTP application (under Oracle RAC with 8 nodes), the array would deliver 6 million IOPS with a latency of 0.4 ms against 275,000 IOPS at 0.9 ms for its best traditional competitor. With SAS Kx, Vexata announces that it can support a bandwidth of 45 GB / s with 240 sessions in parallel (against 4 GB / s and 20 sessions for its best competitor).

This performance can be even higher with the version of the array using Intel Optane disks, the VX-100M. ESG analysts ran a HammerDB test (simulating an OLTP workload) in parallel while an OLAP request consumed 21 GB / s of bandwidth on the array. The result is amazing with a score of 8.5 million IOPS and an average response time reported by Oracle Enterprise Manager of 12 to 30 microseconds.

A very reasonable price and native Fiber Channel support

In addition to these high performances, Vexata highlights several strengths of its architecture. The first is that the bay is a traditional Fiber Channel SAN bay. It is therefore seamlessly integrated into existing storage architectures of large accounts. Unlike emerging competitors in the storage world like Excelero or E8 Storage, there is no need to install specific software components on host servers. And the bay is qualified with most of the big bones of the market, not limited to some Linux distributions.

Second strong point, Vexata pulled the prices. A basic configuration with 4 storage blades is thus proposed from $ 75,000, while the most muscular version with 16 blades and 64 SSDs (approximately 200 TB of raw capacity and 155 TB of useful capacity) is proposed to a public price below $ 500,000. The price of the Optane version of the bay has not been specified.

Oath (AOL + Yahoo) replaces 3Par with Vexata Array

One of the company’s first clients, Oath, the new entity created by the merger of AOL and Yahoo into the US carrier Verizon, agreed to answer our questions about the firm’s bays. The operator acquired several Vexata bays to update its previously motorized architecture with HPE 3Par bays. As Dan Pollack, Oath’s chief storage architect, explained, the VX-100 array was chosen to support the company’s content publishing system, a critical system for the company.

Oath was able to access the first Vexata systems and was able to test them extensively during their development cycle. The firm was able to validate their fault tolerance and performance. “We are submitting the arrays to a comprehensive testing process before acceptance into our architecture and Vexata has passed all our tests,” says Pollack.

By replacing its HPE 3Par arrays with those of Vexata, Oath has not only seen its performance explode, but the firm has also reduced the footprint of storage in its data centers by two.

Asked by LeMagIT about the lack of deduplication in the manufacturer’s arrays, Pollack noted that deduplication is not part of Oath’s required functions. “Our applications are optimized and most of the data they write is already compressed at the base. We do not see any interest in deduplication in our context, “he explained.

A software-defined version in preparation for 2018

Vexata does not intend to stop there. If the firm offers a ready-to-use appliance with the VX-100, it is working hard on a 100% software-based version of its technology. The idea is to be able to offer customers who want the opportunity to implement the technology of the firm on a cluster of NVMe storage servers unmarked. Only small reserve, control nodes will probably continue at least at first, to be delivered by Vexata.