Published: 06 Jun 2013
Over the past decade, compute power has increased nearly 30-fold and networks have increased in speed roughly 100-fold. During that time, hard disk drive (HDD) technology has not significantly increased in either rotational speed or total throughput. Adding a solid-state drive (SSD) tier to an HDD array is like adding a turbocharger to a car engine. But adding an all-SSD array? That's like replacing it with a jet engine.
Part 1: Using all-SSD arrays to turbocharge your data center
Adding more capacity to a storage array increases available performance, as measured byIOPS (I/O operations per second), so storage sizing considerations have to include both capacity and performance.
Crunching the numbers
When comparing the merits of HDD versus SSD arrays, IOPS is a key measurement. Consider that a RAID group with four 400 GB 15 K SAS drives would yield approximately 700 raw IOPS (175 per drive). If an application required 900 IOPS, it would be necessary to add two drives to that RAID group to provide adequate application performance. However, one could not add more data to the RAID group at the same time, because presumably the additional data would require more IOPS to serve it.
In terms of IOPS, 400 GB and 600 GB 15 K SAS drives yield the same raw IOPS. So, a 400 GB HDD has 0.4376 IOPS/GB, whereas as a 600 GB HDD has 0.2917 IOPS/GB. While higher capacity drives deliver a lower cost per GB, they also have lower IOPS/GB. Of course, this is a vast oversimplification of HDD performance, but the complexities are not critical here. The important concept is that IOPS are necessarily tied to capacity. Result: better performance, less efficiency.
The cost of high performance
Solid-state disk has introduced a whole new level of performance to storage systems, and IOPS per GB is a useful ratio for IT managers to consider. Enterprise-class SSD can deliver approximately 100 IOPS/GB, which represents more than 220 times more IOPS/GB than a 400 GB HDD.
But one barrier to SSD adoption has been the perceived high cost. However, while higher capacity drives deliver lower cost per gigabyte, they also have lower IOPS/GB. Array-based SSD can cost up to $40/GB, while enterprise Serial-attached SCSI (SAS) hard drives cost about $5/GB. SSD storage offers 2.5 IOPS per dollar, whereas a 400 GB HDD offers just .15 IOPS per dollar. That translates to 16 times more performance per dollar spent.
The cost per gigabyte of all-SSD arrays may not be shocking; Nimbus Data Systems Inc. sells its all-SSD technology for about $8 to $10 per GB, for example. This price point makes all-SSD very viable for specific workloads.
And a small slice of SSD can make a huge difference in an array's IOPS potential. For example, adding 2 terabytes (TB) of SSD to a 100 TB array -- a 2% increase -- could yield nearly 30% more performance potential. SSD may be expensive on a capacity basis, but remarkably cost-effective on a performance basis.
Enterprises have been adding SSD to their arrays during recent years, boosting performance of HDD arrays with economically feasible amounts of solid-state capacity. But prices have dropped enough as of late to warrant forgoing HDD altogether for a certain subset of applications.
Database acceleration is a common use case for an all-SSD array. This may be particularly attractive for online transaction processing (OLTP) applications where either sustained or bursty I/O activity may exceed the capacity of HDDs to perform. Another favorable workload is virtual desktop infrastructure (VDI), where a small SSD footprint can serve a very large number of devices with consistent speed. A third workload is large-scale server virtualization. The random nature of data access by such a large number of servers into a single storage device might normally require frequent physical hits to the HDD rather than cache. All-SSD obviates the need for slow access to spinning disk. Perhaps the most interesting emerging use case is big data analytics. SSD arrays can deliver enough performance to do real-time analysis on very large quantities of data.