Flash Moving to Mainstream Enterprise Storage?

August 21,20-12, Flash Memory Summit, Santa Clara, CA—Radoslav Danilak from Skyera looked at the changing adoption rates for enterprise flash storage. The critical drivers for change are the physically limited capabilities of the HDDs.

For mass storage, HDDs are one of the best alternatives with low costs per bit and wide availability. Unfortunately, the growing gap between CPU performance and the storage systems is becoming excessive. The CPUs are becoming power limited, but they still far surpass the drives for data volume.

The storage industry has tired to address the performance gap by changing to RAID configurations, but this change causes the costs to double. Short stroking alternate tracks helps but it too adds costs to the drives. The HDDs have moved to 15k RPM motors to increase data rates, but it is not reasonable to consider doubling the speed again. Changing the 15k drive form factor to a 2.5-inch package costs more for the incremental improvements in throughput.

Currently, the only viable alternative to reduce the performance gap is to move to NAND storage. This technology is able to increase performance by a factor of 100, reduce power consumption to 1/10 at a cost increase of over 10x. the SSDs have become the go-to HDD plug-in replacement, but they are hindered by the existing HDD protocols and interfaces.

One way to get better performance is to change the form factor and relocate the storage closer to the CPU. A card-based storage array could be used as a server cache on the server motherboard. Direct attached storage will still have problems since it has to be protocol compatible with the rest of the storage components. By making the DAS an appliance, greater performance is realized through specialized form factors and functions. the problem is that these appliances don’t scale and increase management costs.

If the high performance SSDs move into primary storage, they can overcome most of the I/O performance gap. Unfortunately, the adaptation rate is fairly low, less than 0.3 percent of the enterprise drives are SSDs. A part of the problem is the initial cost. SLC drives cost about $5-7/GB, and the installed cost goes over $20/GB when the various margins stack up from the many vendor layers.

Enhanced MLC sells for about $7-9/ GB and can reduce over provisioning and improve RAID performance. If the user goes with consumer-grade MLC, they can get the SSDs for $2-3/GB and add compression, de-duplication, and intelligent wear management software to get adequate performance and endurance for their apps.

When flash sells for $3/GB in native form, this translates to an $1/GB effective cost. The 19-22nm MLC flash devices have broken the $1 barrier with an actual price around $0.50. The endurance is a function of the physics, which is about 1 percent of the requirements for an enterprise drive. By optimizing the entire storage stack: systems, RAID, controllers, and the flash chips, it is possible to get MLC into the enterprise.

The steps to amplify the life of the devices requires a system that minimizes writes, new DSP algorithms, better ECC, adaptive reads and writes, and attention to the device physics. System architectures like RAID 6 with 3 writes per write help the system, but are bad for SSDs.

Reliability can be enhanced by making appropriate changes in the systems. The architectures, controls, and RAID structures have to address hot spare storage modes. RAID 5 is not sufficient for better reliability, the systems needs to have multiple block-level controls. Adding compression and de-duplication enables enterprise-levels of endurance and reliability for $0.5-0.99 per GB. The effective endurance for a MLC SSD with the enhancements is improves by 100 x, and exceeds that of native SLC drives overall 10 year lifecycle. The same is true for the reliability.

The network becomes the bottleneck, so systems need to move to new architectures. 40 servers going into 2 iSCSI links is ok for HDDs. If the network, storage, and the connections to the top of the rack switch are merged, the network latencies are mostly eliminated.