Enterprises need hyperscale or webscale storage. They look at the likes of Google and Facebook and need to do what they’re doing; processing and storing huge amounts of data with transactions and analysis on the fly. But they lack storage products to do it, and existing enterprise storage isn’t up to the job.
Those are the views of Infinidat, which came out of stealth this year and which believes enterprises are trying to solve the problems of the future datacentre with storage architectures from the past.
Infinidat CTO Brian Carmody said: “What we’re seeing is a competitive advantage by a small number of technically advanced companies – Facebook, Google etc – who are trying to do what they do with incumbent technology. The experience of these apex big data companies demonstrate a failure by the storage industry.”
What Infinidat offers is high capacity (2PB in a 42U rack), high performance (up to 750,000 IOPS, throughput of 12GBps), highly available (99.99999% – seven nines), which it calls mainframe levels of performance and availability but built for enterprise users to handle webscale operations.
It does all this by breaking the mould for enterprise storage in a number of ways. its Infinibox comes with three controllers (in contrast to the standard dual setup) in an active-active-active architecture. These contain DRAM and flash, with all active data held in these two layers, while below that are huge amounts – 480 nearline-SAS HDDs – of spinning disk.
It spreads the workload across all three nodes and out to the massive number of drives. To do that, Inifindat had to throw out the existing Linux SAS drivers and re-write the way the controller nodes handle data to the storage media.
“Each node has a multipath connection to 480 spindles and that is many times more concurrent SAS connections than Linux can handle,” said Carmody.
For now, Infinibox is block access only, but NFS file access is planned “shortly”. Mainframe access will be available by the end of the year and object “shortly after that”, said Carmody.
To achieve such levels of claimed availability Infinidat breaks I/O down into 64kb objects it calls sections with all host I/O being dealt with to and from cache. This is then spread out across the HDDs in a 16-way stripe with parity, based on the RAID 6 model but dubbed Infiniraid. Data is given an “activity vector score” to rank it on whether hot or cold, sequential or random etc. This helps with pre- and de-staging data between tiers and also for rebuilds.
“At hyperscale n+1 is not sufficient,” said Carmody. “We have three of everything. It’s what hyperscale in the enterprise will look like.”
That’s an intriguing claim. Watch this space.