There is a lot of confusion today when I meet with customers on what constitutes enterprise flash. Some array vendors claim eMLC (Enterprise MLC) flash is an absolute necessity for a reliable all flash array. This is a myth but I am afraid I can take my share of the blame.
Five years ago, I made the switch from Apple to Sun with the goal of bringing the technology disruption of flash to the enterprise. It was clear there weren’t many options for building flash
solutions for the data center. There were a few exotic enterprise controllers that were extremely expensive. The consumer SSD controllers weren’t capable of meeting the demands of heavy workloads in conjunction with the NAND available at the time. The strategy was simple. Instead of using an enterprise controller to make NAND flash work, make higher quality NAND flash to pair with a consumer controller to meet the challenges of enterprise workloads at a significantly cheaper cost. A couple meetings with NAND flash vendors later, enterprise grade NAND flash was born. There were four fundamental changes needed to make NAND flash ready for the data center.
- Slow down NAND flash tries to program data as fast as possible due to customer requirements for more performance. This destroys the precious oxide layer protecting the electrons in the cell storing data causing the NAND to wear out more quickly. Slowing down programing protects this oxide layer enabling longer life. Several vendors were surprised when NAND flash that would fail at 300,000 program and erase cycles was capable of 1,000,000 cycles without failing by simply slowing down.
- 24×7 Optimized At the time, the NAND Flash was designed for camera cards and music players that didn’t use the flash very often. Heavy workloads in servers stressed the flash causing failures in control logic that had never been seen before. The first enterprise grade NAND needed a redesign adding robustness to this logic to meet demands of 7×24 continuous workload.
- Change the Spec Flash specifications on reliability and endurance don’t specify the actual failure characteristics of the flash but rather are used for warranty and failure rate negotiations with large customers. Enterprise flash at the time didn’t require long-term storage of data since most applications were for caching. Flash vendors were happy to lower the worse case data retention from one year to three months in order to increase the number of write cycles in the specification.
- Cherry Pick Intel and others have built a business model around taking the best wafers and die from their factories and selling them as fastest processors for top dollar, while the remaining wafers and die are sold for a lower price and marked as a slower processor. The same is true for NAND flash. The best parts go to enterprise grade NAND flash, the good parts go to large customers and the lower end parts end up in those USB memory sticks you find at the check out counter at your local drug store.
The flash vendors were happy to create the enterprise grade flash because it meant a 20%-40% price premium over standard flash for very little work. The strategy was extremely successful and Sun’s enterprise grade NAND was showcased in Oracle Exadata. It was the right solution for the time. But today, robust logic for continuous operation and slow programing features are built into all NAND flash. Vendors are still just cherry picking when they want to turn a standard part into an enterprise grade one. High-priced enterprise flash isn’t needed today because the controller technology has finally caught up to the market.
Five years of research and development in controller-based solutions have enabled standard flash to perform better and more reliable than enterprise grade flash with consumer controllers. These controller vendors worked in close collaboration with the NAND vendors and together made breakthroughs in understanding how flash fails and techniques for increasing reliability and longevity. For example, Apple acquiring Anobit, Hynix acquiring Link A Media Devices, Micron acquiring Storage Genetics and Seagate partnering with Densbit all demonstrate that controller technology is key to making commodity flash solutions.
Delivering an all-flash array at the same cost as a performance disk is the mission of Pure Storage and standard MLC NAND flash plays a significant part in achieving that cost breakthrough. Pure has an innovative strategy hardware/software solution for bring standard MLC NAND to the enterprise market.
First, Pure has a strategic partnership with and investment from Samsung, the largest manufacturer of NAND flash. We leverage the Samsung Data Center MLC SSD. This product pairs a commodity laptop SSD with special firmware optimized for server and storage applications as well as special features for our Purity Software. The Samsung internally developed SSD controller used in this drive is a derivative of the controller used in the Samsung enterprise SLC SSD shipping in many storage arrays today. This SSD controller/standard MLC combination from Samsung is significantly cheaper than any eMLC SSD solution we could find on the market today, and yet is reliable for enterprise grade workloads.
Second, Pure created a technology called FlashCare™. This important layer in our Purity software globally manages the SSDs to not only perform in the best way possible but also to increase reliability by eliminating unnecessary operations done by the SSD and improving longevity. FlashCare™ accomplishes this global wear leveling and micro-scheduling I/O to insure requests are aligned and tightly controlled to avoid overlapping scheduling of reads and writes.
We aren’t alone in using standard MLC in the enterprise. STEC CellCare and Smart Storage Systems Guardian offer controller technology enabling standard MLC in enterprise SSDs that eliminate the unnecessary margin paid for enterprise MLC. Soon others will implement similar technology that will bring an end to enterprise grade flash and enable a new era of controller driven flash solutions with better reliability and lower cost using standard flash. Long live flash.