Andy Walls of IBM talks about NAND Flash for Enterprise Applications

Just got back from a morning spent at the Flash Memory Summit. The last talk I listened to was the pre-lunch keynote from IBM’s Andy Walls, a Distinguished Engineer who has worked at IBM for 29 years and has lots to say about enterprise storage. Walls started his keynote by discussing the 4-legged stool for a great SSD strategy. The four stool legs are:

1. Enable Enterprise MLC (multi-level cell) NAND Flash. You do this, said Walls, through write-mitigation techniques and large-capacity Flash arrays that allow overprovisioning.


2. Flexible packaging alternatives. No one size Flash package fits all. You need NAND Flash configured in SSDs, in board-level modules, and in shelves. The packaging of choice depends on the application and use case.


3. SSD-optimized infrastructure. We’ve spent 50 years optimizing operating systems, middleware, and applications for HDDs with 5-10 msec response times. It will take some amount of time for us to optimize all of that software for the faster response times we get from various NAND Flash storage devices.


4. Differentiating storage software (what Walls calls “the real differentiator”). Vendors need to help customers figure out what data to put on SSDs either through manual placement methods or through automatic tiering.

SSDs are game changers for servers, said Walls. They provide far more IOPS than HDDs; they reduce access delays; they have reduced I/O wait times; they deliver many more IOPS for fewer Watts of power (30K IOPS at 6W for SSDs versus 300 IOPS for 9W with an HDD).

(Note: Walls cautioned that all SSD IOPS are not the same. Write IOPS are far slower than read IOPS, for example.)

Finally, SSDs have the potential for high reliability said Walls. However, to get that reliability, we need to deal with the wearout failure mechanisms inherent in MLC NAND Flash. “We know how to build reliable electronic assemblies,” said Walls. “We’ve known how for decades.” We just need to figure out how to improve MLC NAND Flash endurance, he said.

When we do figure out how to deal with the endurance problem, there are three primary data-center uses for MLC NAND Flash:

• Put hot data on the SSD using manual placement, assisted placement, or automated placement tools.


• Temporary data placement. Storage applications including data warehousing, paging, and caching all deal with temporary data and all benefit from an SSD’s improved access speeds.


• Fit all of an application’s data onto SSDs. Walls noted that at least half of all databases are smaller than 4Tbytes. In data-center applications, that’s a storage capacity that can be economically achieved with SSDs alone. Workload-optimized servers especially can benefit from SSDs to really maximize performance and in some cases, where processor loading is only 10% or so due to disk-related bottlenecks, a data center can reduce server count by 10x by switching certain servers to all-SSD storage.

As a result of these radical changes in bottlenecks caused by fast SSDs relative to slow HDDs, Walls sees a return to direct-attached storage rather than the currently prevailing, years-old trend towards SANs. All sorts of servers will benefit from this major architectural upheaval. In particular, Walls pointed out that analytics engines--the servers that sift through databases to find trends in financial markets, retail shopping patterns, and criminal activity--can especially benefit from the use of SSDs for storage.