Smart Keynote

August 21,20-12, Flash Memory Summit, Santa Clara, CA—John Scaramuzzo from Smart Storage Systems talked about the internal and external drivers for the storage industry. His talk was framed around the question, “if time is money, what is endurance related to?”

As a background, the exponentially growing volume of data is increasing the complexity of the connected world, increasing the number of channels for data creation and consumption. To illustrate, 72 hours of video are uploaded to YouTube every minute. The traffic on the Web is equivalent to 7 M DVDs per hour. Total Internet traffic is expected to exceed 667 exabytes by ’15.

The drivers for storage are: in memory computing, DRAM being displaced by flash, and the need for low latency storage. The data center is becoming part of the cloud and the number of apps is increasing geometrically. The performance needs for storage require an operating load from 0-100 percent read-write mix.

For HDDs, the endurance requirements and the workloads measured in drive writes per day—the equivalent data traffic equal to the full capacity of the drive—ranges from 1-20. NAND in its many incarnations of MLC at 0.4 drive writes per day, eMLC at 7 dw/d, and SLC at 20 dw/d can meet the workload requirements as HDD replacements. The challenge is to match the stated endurance to the workload. In many cases, the workload is an under sampled function, implying the actual workload may vary considerably from the stated requirements.

This loss of resolution for a parameter means that some significant percentages of the drives are over provisioned relative to the requirements. For example, a file server with a 90 percent read cycle is specified at 10 dw/d. MCL storage is usually specified at 0.1 dw/d, so this storage type seemingly needs to be replaces often. eMLC is specified at over 7 dw/d.

This level of workload is actually overkill for the situation, since only 10 percent of the drive accesses are writes, so the real number is close to 1 dw/d. The eMLC drives are about 3 times the cost of a MLC drive. At a PB of storage, the cost differential is reduced to about two times. By adding more error correction and better wear leveling software, the MLC drives can handle the workload, so an IT manager can safely buy MLC drives and expect sufficient endurance for the application.

On-line transaction processing usually calls for 10 dw/d and represents a 70 percent read cycle. An eMLC device at 7 dw/d seems incapable of meeting the workload requirements. An SLC drive with a dw/d of 20 can easily meet the needs, but is going to increase the total cost of operations by over 18 percent over a 5 year operating span at a 1 PB capacity. In addition, the system will require more work to configure the tiering software and doing basic management tasks.

By using software that minimized writes, the user can consider the eMLC as a viable alternative. The issue for the storage vendors is to improve the underlying DNA of the MLC flash devices. These devices need to improve reliability and endurance to meet the stiffer requirements of a high-volume workload. Although endurance enhanced MLC drives cost more than the plain versions, the total cost of operations greatly exceeds the cost savings of 82 percent over a SLC drive.

Endurance has to be matched to the workload. Online transaction processing can benefit by using eMLC rather than eeMLC drives. The acquisition costs are lower and the total cost of operations is twice as good. Loss costs money. By grouping apps into endurance points, the system operator can configure storage to meet the realistic needs, and save acquisition and operating costs.