Storage Systems - SQL Server 2012 Internals and Troubleshooting - page 105

is manufactured using multiple stacked cells that share a single transistor. This enables MLC to be
manufactured more economically, but it has the drawback of lowering performance and increasing
the potential failure rate.
Enterprise MLC (eMLC) is a newer generation of l ash technology that increases the reliability and
performance of multi-level l ash. Before choosing a type of l ash, it is important to understand the
number of reads and writes that your l ash will be subjected to. The acts of writing and erasing from
l ash cells degrade the cells over time. Flash devices are built with robust error correcting and check-
ing, but they will fail due to heavy write cycles. Ensure that the l ash technology you implement will
survive your application performance requirements. Unfortunately, l ash devices are not a one-size-
i ts-all solution. If your application generates an extremely high number of random I/O requests,
l ash will probably perform well. Conversely, if your application creates a few large requests, such as
a data warehouse application, then you won't necessarily see a great benei t.
Several manufacturers sell a l ash-based PCI express NAND l ash device. These cards offer extremely
low latency at high I/O rates. A single card will respond in tens of microseconds and generate hun-
dreds of thousands of IOPS. Contrast this with a shared array that responds in hundreds of micro-
seconds, or even milliseconds. If your application can generate this type of I/O load, these cards can
greatly increase your potential performance. Not only can the card sustain hundreds of thousands of
I/Os, each is returned much faster. This can relieve many SQL blocking issues.
We stress that your application must be able to generate appropriate I/O because we have seen many
instances in which customers have installed l ash as a panacea only to be disappointed with low per-
formance increases. If a given application is hampered by poor design, throwing money at it in the
form of advanced technology will not always i x the problem!
There are now hybrid PCI express-based solutions that combine server software, the PCI express
l ash card, and shared storage arrays. These systems monitor I/O access patterns. If a given work-
load is deemed appropriate, data will be stored and accessed on the PCI express l ash card. To main-
tain data integrity, the data is also stored on the shared storage array. This hybrid approach is useful
for extremely large datasets that simply won't i t on a series of server l ash cards. In addition, SAN
features such as replication can be blended with new technology.
Many shared storage arrays offer l ash solutions that increase array cache. These systems work
just like the PCI express hybrid solution, except the l ash is stored inside the shared storage array.
Appropriate data is migrated to the l ash storage, thereby increasing its performance. As stated
before, if the access pattern is deemed not appropriate by the array, data will not be moved. Heavy
write bursts are one example. A massive checkpoint that attempts to write 30,000 IOPS will prob-
ably never be promoted to l ash because the accessed data changes every minute!
Shared storage arrays blend tiering and automated tiering with l ash drives. When you consider most
databases, only a subset of the data is in use at any given time. In an OLTP system, you care about
the data that is newly written and the data you need to access over a short window of time to derive
metrics. Once a sales quarter or year has passed, this data is basically an archive.
Some DBAs migrate this archive data. Automated tiering offers a low-overhead system that actively
monitors data use. Active data is promoted to a l ash storage tier; moderately accessed data is
migrated to a FC or SAS tier; and archival data is automatically stored on high-capacity SATA.