Startup Sees Potential For Long-Distance RDMA On Horizon - The Next Platform
Published in The Next Platform, January 10, 2019
Sharing data via remote direct memory access (RDMA) has typically been a local affair, restricted to a single server or tightly bound clusters of servers. But Silicon Valley startup Vcinity has something else in mind.
RDMA has turned out to be a highly efficient way of exchanging data between processors in a single computer or over a local area network (LAN). By bypassing the CPU and going directly to DRAM, computations are able to take place in parallel with memory reads and writes, which has made it extremely popular in high performance computing and other throughput-demanding environments. InfiniBand, RoCE (RDMA over Converged Ethernet), Omni-Path, and iWARP all support the technology natively.
A recent stealth mode emergent, Vcinity is looking to bring RDMA technology to long-distance networks – wide area and metropolitan area networks – using its fabric extender solution that can be hooked into conventional servers and storage. In a nutshell, the Vcinity hardware leverages the high performance capabilities of InfiniBand and RoCE to create a network fabric that can move data over thousands of kilometers with low latencies and high bandwidth. And since the fabric uses InfiniBand or Ethernet as the common denominator, from an application point of view, the technology is, in theory, seamless.
For organizations whose digital data is geographically dispersed, the Vcinity fabric targets the ability to crunch on remote data in-place without the need to copy whole files back and forth across datacenters. These days, that kind of capability is extremely useful to a lot of businesses and government entities, including media companies, oil and gas firms, healthcare organizations, defense agencies, and national research labs – basically any organization with data sprawl challenges.
Vcinity has implemented this with its own network extending hardware that makes remote data look like it’s local, or at least look like it’s part of a unified system. Not only does this provide a more convenient experience for users, it avoids both the time penalty of copying entire files and the complexity of keeping those copies coherent.
Constructing the illusion of all-local data is possible because Vcinity has developed a global data fabric that exhibits the performance and RDMA characteristics of InfiniBand and RoCE upon which it is based. The illusion can be maintained because, according to Vcinity, their solution makes it possible for remote data to be accessed in “near-real-time.” Although the company isn’t sharing a lot of details on how it works, one can assume they are taking advantage of some combination of aggressive port buffering, advanced flow control, and a healthy dose of data parallelization.
Of course, in some cases, you actually do want to copy your data to different locations, namely in those instances where files need to be moved or replicated. Processing data in-place is all well and good, but organizations are always going to need file back-ups and disaster recovery. Vcinity says their technology can do this as well, and since it’s using the same RDMA fabric extension technology, the company says it can perform these transfers at extremely high speeds.
In fact, they claim they can deliver better than 95 percent utilization of available WAN bandwidth in many cases. If that’s the case in production, that is significantly better than what standard TCP/IP can deliver, which is typically in the neighborhood 35 percent. It could also be a good deal more convenient than shipping disks or tapes cross-country, which, believe or not, people still do. To prove its point, Vcinity demonstrated a one petabyte data transfer from California to Maryland in under 24 hours.
Being a startup that emerged from stealth just four months ago, there is bound to be some skepticism about the company’s claims. Vcinity has documented a handful of use cases from some early customers and commercial partners. More comforting is the fact that some of the underlying technology has a much longer track record. The company’s roots are actually in Bay Microsystems, a 20-year-old company Vcinity acquired in July 2018. Bay provided long-haul InfiniBand, primarily for the defense industry using this same RDMA fabric approach.
According to Vcinity CEO Harry Carr, who also headed Bay Microsystems prior to the acquisition, the original product set was powered by custom ASICs and was narrowly targeted to military and security applications running in HPC-type environments. Vcinity generalized the design, switching to off-the-shelf FGPAs (from both Intel and Xilinx) and adding support for RoCE in order to address a wider audience. Vcinity will soon be offering a software-only version, promising availability later this quarter.
At this point, Vcinity’s most basic piece of hardware is its Radical X platform, a 1U box that supports up to 600 Gbps of non-blocking switching capacity. It provides multiple RoCE or InfiniBand links at speeds of up to 100 Gbps. According to the spec sheet, at 100 Gbps speeds, it can deal with up to 250 ms of latency. Radical X switches will most likely end up being integrated into third-party database and file system appliances that are required to support storage dispersed across multiple datacenters.
Radical X also is the building block for Ultimate X, Vcinity’s stand-alone appliance that is able to hook into an enterprise network and provide its patented fabric extension or act as a unified data exchange. The appliance can talk to NFS storage, but also supports HPC parallel file systems like IBM Spectrum, Lustre, and BeeGFS.
Even though Carr thinks there is a big market for long-distance RDMA in enterprise and cloud computing environments, HPC users may end up being the earliest adopters. Given the familiarity of these technologies in the high performance computing community and its own needs for big data, that makes a certain amount of sense. According to Carr there is a current project with two HPC sites in Southeast Asia, as well as some feelers from a number of other potential HPC customers.
Although Vcinity does offer direct sales, Carr and company are looking to do most of their business via channel partners, primarily OEM or system integrators that can incorporate the fabric extension hardware and software into existing server and storage products. It’s worth noting that two of those partners are HPC specialists Cray and Penguin Computing. As of now, none of the big five system OEMs (HPE, Dell EMC, Lenovo, IBM, and Cisco) have signed up, so Vcinity has its work cut out for it.
Like any fledgling Silicon Valley startup, the company is on the lookout for investors. Carr says they are currently funded by angel investors and have not taken any venture capital. Vcinity is seeking its A Round of funding, targeted primarily at strategic investors.
Written by Michael Feldman, Senior Editor, The Next Platform