A3Cube turns RAM up to 11 with FORTISSIMO kit
In-memory fabric comes to to HPC I/O project
A3Cube has fleshed out another part of its plans for high performance interconnect, adding an I/O access system to the RONNIE data plane technology it launched in February 2014.
As explained to El Reg when it hit, RONNIE provides direct, switchless connection between nodes using PCIe NICs, claiming around 900 nanosecond latency between CPU, memory and I/O.
The next piece of the puzzle is its newly-minted Fortissimo Foundation, which the outfit says targets “large scale-out applications” such as NoSQL databases, Hadoop, Object Storage, and NAS.
The foundation is an A3Cube-made Linux-based “in-memory fabric”: there's the “megafiler”, which aggregates memory and storage over DMA; and a management engine.
“Local-CPU to remote device access times of less than 0.8usec are achieved by Fortissimo with RONNIE Express, significantly improving access times compared to traditional storage configurations and enabling storage performance to scale linearly across multiple nodes,” the company's release claims (PDF, here).
Stripping the gobbledygook out of the release, the key thing seems to be that it's an in-memory scale-out. Fortissimo creates a global name space that can aggregate all the resources of a data centre cluster to behave as if it were local (which is where RONNIE's 0.8 microsecond latency is important – so that the system doesn't notice the time it takes to get to non-local memory).
The existing enterprise storage environment, A3Cube claims, can get high IOPS performance, DRAM scalability to petabytes, and communications without involving the OS kernel. It claims to work with “any combination of memory, flash, SSD, hard disk, CPU or data centre I/O fabric” and comes with the remarkable claim that it enables a 16:1 consolidation of cluster server sizes (YMMV). ®