There’s a new kind of flash-enabled application architecture coming, and Wikibon co-founder David Floyer (right) expects it to do a lot more than take a couple of ticks off the clock. He thinks it will revolutionize the speed at which organizations deploy real-time analytics.

The key components of the new architecture are high-speed DRAM memory, flash storage and a new kind of backplane technology based on the Peripheral Component Interconnect Express (PCIe) bus. The processors in the cluster share flash storage through an enhanced PCIe switch that provides shared access to a large number of dense flash modules for ultra-low latency, enabling parallel persistent reads and writes.

One version of such a bus is built by DSSD, Inc. a startup EMC acquired last year for its technology for pooling server-based flash storage for high-performance data access. Other companies will enter the market with similar technologies, however.

What makes the Flash-as-Memory-Extension (FaME) architecture so revolutionary is that it all but eliminates latency and jitter – or variance delays – that plague conventional disk-based clusters and even all-flash arrays. It thus achieves a breakthrough that previous enterprise performance clustered systems couldn’t: delivering high performance, scalability and availability at reasonable cost.

“The difference with FaME is that instead of having the network outside the processors, it’s now inside,” Floyer said in an interview with SiliconANGLE. “The network is very fast using a board sitting in the middle of the cluster. So if I want a piece of data from a particular module it will switch that data to me very quickly, in microseconds rather than milleseconds.” A microsecond is one one-thousandth of a millisecond.

The use of high-speed DRAM ensures maximum throughput, and the pooled flash storage acts effectively as a feeder for main memory, as well as a backup in case of a processor failure. The critical innovation was the addition of a fast, low latency, high bandwidth switch in the middle of the cluster. “That is where the bottleneck is today,” Floyer said.

In other words, the architecture makes sure that processes are never tapping their electronic feet waiting for data to be fetched from storage. This latency problem tends to have a cascade effect in highly integrated applications because none of the processes can move forward until all of them have the data they need. A bottleneck at any node in the cluster can bring the entire cluster to a standstill.

“In a clustered architecture you typically have a very large amount of waiting going on,” Floyer said, noting that networked architectures typically work at 10% of capacity, compared to up to 90% for mainframes.

FaME can potentially yield blistering performance, loading a one-terabyte database in minutes, compared to hours for a conventional disk-based cluster. More important is that it can enable an organization to conduct analytics in near real-time, since FaME can integrate the analytics into the operational systems and automate decisions. Flash memory supports low-overhead data sharing in ways that are impossible using traditional magnetic or SSD disks.

That’s what’s revolutionary about the architecture. “Applications in this environment can process thousands of times more data than they currently can,” Floyer said. “Instead of analytic computing making one or two people in the company think smart, it can make an entire company act smart.”

Floyer’s extended report is available at Wikibon Premium.