If startup Symbolic IO is correct, the way storage systems are designed could be set to undergo a big change. Before...
that happens, though, the company has its work cut out educating enterprise customers how its in-memory storage technology operates.
Following more than four years of development, and with nearly $13 million in funding, the Holmdel, N.J., vendor this week previewed its branded Intensified RAM Intelligent Server (IRIS). By changing how binary bits are processed, Symbolic IO claims IRIS can shrink the size of data stored in RAM, rehydrate it quickly and preserve the data intact.
IRIS consists of a commodity 2U server based on Intel's Haswell processor, along with system BIOS modified with intelligence to recognize nonvolatile memory as persistent storage. Founder and CEO Brian Ignomirello said Symbolic IO will start shipments this year of three branded configurations: IRIS Compute, IRIS Vault and IRIS Store.
The company claims to have created a method for in-memory storage that rewrites the instruction sets used for encoding binary data in RAM. Ignomirello described it as computational-defined storage. That description refers to the technique Symbolic IO uses to issue commands to create and recreate data as users request it.
Ignomirello served as CTO of Hewlett-Packard's global storage team and also worked at NetApp and EMC before launching Symbolic IO in 2011. The focus of Symbolic IO was to decouple computational tasks from dependency on the underlying storage media. He said IRIS lets customers take advantage of idle server RAM and avoid the cost of buying flash as an acceleration tier.
"We move the entire storage process directly in front of the processor itself. In doing so, we created a way to overcome traditional file limitations," Ignomirello said.
IRIS in-memory storage creates and recreates data, but doesn't store it
Symbolic IO's patented technique receives a block of unencoded raw data, deconstructs it into a plurality of data vectors and maps the data vectors to a bit marker table stored in RAM. This step produces an encoded representation of the raw data.
To decode data, Symbolic IO retrieves the mapped bit marker from memory and combines it with data vectors and other blocks to produce a composite block of decoded raw data.
To recreate the data, IRIS recalls it based on queries and applies proprietary algorithms in the Symbolic Conversion Engine, a thin OS that acts as a bootstrap to load and control all binary compute fields.
Brian IgnomirelloCEO, Symbolic IO
"We don't store the data. We give instruction sets to materialize and dematerialize data on the fly. All of this is done in a noncompressive fashion and is computational-bound, so it happens in real time," Ignomirello said.
Scott Sinclair, a senior analyst with Enterprise Strategy Group Inc., in Milford, Mass., said Symbolic IO takes a contrarian approach to legacy storage architectures.
"They've figured out a way to represent a larger amount of data within a smaller footprint of bits," Sinclair said. "Rather than adapt a data layout architecture designed for disk, they're going in the opposite direction. They believe the future will be dominated by in-memory storage and are focused on how processors talk to memory, rather than how applications talk to disk."
IRIS speed opens beta customer's eyes
Telecommunications firm MetTel in New York is beta testing the IRIS Compute server. Will Prince, MetTel's vice president of operations, said his IT team needed several demonstrations to fully grasp the nuances of Symbolic IO's in-memory storage approach.
"They've taken a traditional x86 server build and modified it," Prince said. "Instead of using solid-state drives or flash, they take a certain percentage of the RAM slots and use those as [the equivalent of] hard disk. Each of the DIMMs has a capacitor backing it up. For all intents and purposes, they've made the RAM as hardened as the hard drive itself."
MetTel ran a virtualized replica of its SQL Server production environment on IRIS, consuming four processors and 4 GB of RAM. The physical SQL Server workload used 12 processors and 64 GB of RAM, "but the Symbolic box was about four times faster, because it reads and writes directly from disk," Prince said.
Prince plans further testing to insert IRIS in an always-on SQL Server cluster. "If it behaves nice and neat, I will make it the primary server for that cluster," Prince said.
The base IRIS Compute starts at roughly $80,000. Pricing for the Vault and Store models has not been set.
IRIS Store is the scale-out storage model. It can be outfitted with up to 380 TB of raw storage by inserting traditional SSDs, PCIe cards or NVMe drives for creating tiered storage. Effective capacity is expected to be 1.8 PB in 2U. Up to 20 DIMM slots can be provisioned as IRIS Store modules, with four slots provisioned as RAM modules. Ignomirello said the vendor is aiming to pack 128 GB of RAM on an IRIS Store module.
IRIS Vault adds SymCE software and a multimodal ASIC that manages physical security, backups, clones and replication. SymCE includes functionality known as Blink, which uses an embedded NVMe card as a backup engine to capture images of physical IT infrastructure for instantaneous recovery.
A geolocation feature enables IRIS to know if the box has been moved, and a trauma sensor system detects if the machine has been dropped or abused, including the angle and force of impact, time of day and other information. If the device is lost or stolen, IRIS Vault will shred data or render it unusable to anyone but the data owner.
ESG's Sinclair said IRIS provides compelling features, but Symbolic IO needs to fine-tune its message to make inroads in enterprise storage.
"Their greatest opportunity is probably the same as their greatest challenge," he said. "They're building a storage architecture that's designed for memory, not disk, and the opportunity lays in all the benefits that stem from that. The challenge is people don't immediately get how it works. That's the bridge they have to cross as a company."
Remember these tips for recovering in-memory databases
Time is in-memory database recovery's top challenge
Why DIMM is the bright future of in-memory storage