The Past
Moore’s Law enabled the CPU to process data faster than the hard disk drive could get the data to the CPU. Because getting data to the CPU was the bottleneck, the database that solved that bottleneck would have a performance advantage.
The shared-disk database had two glaring deficiencies: (1) it required expensive shared storage; (2) running over Fast Ethernet that most data it could get to the CPU was 12.5 MB/second.
Compare this to the shared-nothing database, which: (1) runs on inexpensive commodity servers; (2) delivers up to 64 MB/second of data to the CPU from the local hard disk. Shared-nothing had the advantages of performance and price! It is no surprise that it grew to dominate the database market.
The Present
A lot has changed. SANs, NAS and clustered storage, leveraging fast interconnects like fiber channel, Infiniband, Dolphin Express and others, deliver up to 400 MB/second of data to the CPU, 6-TIMES the performance of the local disk. But these high-end storage devices are still too high for most users. That is changing now.
Cloud computing is based on virtualization and the sharing the IT infrastructure across large numbers of users in a multi-tenant model. By amortizing both the capital expenses and the operating expenses across its users, cloud companies are rapidly commoditizing high-end computing resources. The net result is that users can run their software on a high-end managed cloud with a super fast SAN or NAS for less than running their application on their own commodity servers. As a result shared-nothing databases lost both of their advantages in performance and price.
The Future
Using a shared-nothing database introduces significant complexity for enterprise systems. As you scale to multiple servers, you must partition your data, set-up slaves and replication, deal with slave promotion, address the issues of data inconsistency, data shipping and more. The shared-disk architecture eliminates all of these headaches.
Until recently, these headaches were simply the price you paid in exchange for shared-nothing’s price/performance advantage. Now shared-nothing’s price/performance advantage is gone. Provisioned cloud storage is the new price/performance leader. Now that you are using shared storage there is no longer a rationale to accept all of the deficiencies of shared-nothing. If you are using shared storage, a shared-disk database is simply more compatible.
Shared-disk databases are inherently easier to manage and more cost-effective for cloud companies. Virtualization is a powerful driver of the underlying economics of cloud computing. It allows the cloud company to efficiently host multiple companies by tapping into a pool of compute resources, instead of dedicating specific compute resources to specific companies. For example, a single server may adress the computing demands of 10 or more different companies throughout the day. This is far more efficient than dedicating that machine to a single company. Unfortunately, the shared-nothing database is tightly linked with a physical computer where the data is stored. The shared-disk database, on the other hand, separates the computing form the data, so it can be easily virtualized and distributed over a multi-tenant computing infrastructure. The net result for cloud companies is that because shared-disk databases support virtualization, they reduce the cloud company’s costs and increase profits.
For the reasons defined above, the cloud will usher in the age of the shared-disk database.
