Start-up profile: TidalScale, building an inverse hypervisor for scale-up servers

TidalScale, a start-up based in Campbell, California, is on a mission to build the world's largest virtual servers based on Intel x86 commodity hardware.

The company's "inverse" hypervisor combines multiple physical servers (including their associated CPUs, memory storage and network) into one or more large software-defined virtual servers. This is the inverse equivalent of VMware because a rack of physical servers are virtualized as though it were one. The concept is to scale-up a virtual server instance to handle Big Data workloads without making changes to applications or operating systems.

Why use another hypervisor to create a bigger server? Doesn’t Moore’s Law already deliver more powerful processors over time? And why not just provision a large number of individual servers from a Cloud IaaS vendor? The answers here would be (1) very large in-memory datasets (2) Moore’s law is not keeping pace with rising workloads demands (3) too costly and too limiting, especially since public cloud operators tend to limit the memory size of bare metal servers to 2TB and because in load balancing a workload there is a tendency to provision to more resources than necessary.

The TidalScale story

TidalScale was founded in 2012 by Dr. Ike Nassi, an Adjunct Professor of Computer Science at UC Santa Cruz, who has been involved in many tech developments including as Chief Scientist at SAP when the category of in-memory databases was established. He also was involved in 3 previous start-ups: Encore Computer, a pioneer in symmetric multiprocessors; InfoGear Technology, which developed Internet appliances and services; and Firetide, a wireless mesh networking company.

The technical team also includes Dr. David Reed as Chief Scientist, who holds many patents along with four degrees from MIT in EE and CS including his PhD. Reed's contributions to the networking field include work on the original Internet protocol design team. His architectural contributions included the UDP protocol design, the “slash” in TCP/IP, and formulation of the End-to-End Argument as its primary protocol design principle. Later, he went on to become  Chief Scientist at Lotus Development Corporation, an HO Fellow, and an SVP at SAO Reseach.

On the management side, TidalScale is headed by Gary Smerdon, who previously was the EVP & Chief Strategy Officer of Fusion-io, the devel.oper of flash-based PCIe hardware and software solutions that was ultimately acquired by SanDisk in 2014 for $1.3 billion. Before that, Smerdon was SVP and GM of the Accelerated Solutions Division at LSI, an internal startup that he founded. Smerdon also held executive positions at Greenfield Networks (acquired by Cisco), Tarari (acquired by LSI), Marvell, and AMD.

TidalScale, which first began shipping in 2016, aggregates all the resources (memory, cores, storage and bandwidth) of low-cost, high-performance, 2-socket Intel x86 servers into one or more Software Defined Servers. This accomplished by running a TidalScale HyperKernel on the physical server and a "WaveRunner" control plane and management console to orchestrate the spinning up or spinning down of virtualized servers. The HyperKernal instance on each physical server communicates with other HyperKernal over the Ethernet network, which essentially functions as a combined memory and I/O bus. Thus memory performance will be determined by the latency and throughput of the Ethernet connection. Still, for applications such as very large in-memory databases, a TidalScale software-defined server consisting of five physical nodes each with 128GB of DRAM, will be better than a single server with 128GB of DRAM if the memory required exceeds 128GB and a secondary SSD must also be employed. This is because DRAM performance is roughly 1000X that of flash memory.

Software-Defined Servers can be configured with dozens or even hundreds of processor cores, tens of terabytes of memory, and as much storage and networking I/O as needed. The configuration of servers can be automatically right-sized to the workload. TidalScale allows Docker containers and container management platforms (Kubernetes) to run on top. For instance, TidalScale could be used to deploy a single Linux instance with 15TB of DRAM and up to 400 cores by leveraging dozens of servers in a cloud data centre.

As mentioned above, TidalScale's paradigm scale-up paradigm on commodity servers should be especially relevant to in-memory databases, such as SAP HANA. The company says it can configure up to 64TB of in-memory performance on 2-socket Intel x86 servers. Currently, cloud customers can TidalScale to on standard servers available on IBM BlueMix, OrionVM’s Wholesale Cloud Platform, and Oracle Cloud Infrastructure, with virtual systems ranging from dozens to hundreds of cores and featuring up to 30TB or more of memory. Natural allies then would include any company in that database ecosystem. Because TidalScale was exhibiting at the Open Compute Project Summit, it reasonable to assume that it sees the hyperscale cloud companies also as potential customers.

TidalScale has received a number of awards, including being named a Gartner Cool Vendor, an IDC Innovator for 2017, a Red Herring Top 100 North America recipient for 2017. Another milestone occured in November 2016 when Infosys made an equity investment in TidalScale. Financial terms were not disclosed. Crunchbase says TidalScale has gone through several rounds of venture funding, raising at least $11.8 million, probably more.

In the broader context of software-defined data centres, the need for scale-up servers will certainly be just as important as scale-up storage. Many start-ups have pursued the JBOF (just a bunch of flash) storage array opportunity, and some of these companies were acquired at nice premiums and other completed IPOs. The software-defined server space likely won't have as many start-up entrants, giving this company a better chance at driving its inverse hypervisor paradigm forward.