IT organizations often default to tier one storage because users demand high performance. However, tier one is the most expensive solution and depends on high spindle counts for aggregate IOPS. Combining SSDs with tier two or even tier three HDDs can deliver more IOPS than tier one HDDs at a lower price point.
A substantial number of IT organizations remain committed to tier one HDDs as a comfort zone for delivering optimum performance. For the purposes of this discussion, we'll define these devices as 300 GB or less, 15K-rpm Fibre Channel (FC) HDDs. The key advantages of a tier one strategy are a simplified architecture and insulation from any performance criticism. It's also relatively easy to calculate IOPS by adding up the IOPS per device. However, in many cases, storage administrators must spread LUNs across large numbers of spindles in order to get the necessary aggregate IOPS for specific applications. All too frequently, this results in very low utilization -- as low as 20% -- which increases the effective cost per gigabyte stored by 200% to 400%.
Adding an SSD cache tier to an array can substantially increase the performance of the array, but at a general-purpose level. IOPS requirements frequently follow the 80/20 rule, where 20% of the applications consume 80% of the IOPS. With an array-level SSD deployment, applications that do not need acceleration may consume valuable IOPS that could go to higher-value applications. Now, vendors are increasing the flexibility of SSD with "hybrid tiers" that include both HDDs and SSDs in the same volume. This targets the SSD to applications that need it the most. The huge performance gain with SSDs can allow IT organizations to migrate from tier one HDDs to tier two or even tier three and still deliver the necessary IOPS to specific apps. Here's how the math works. First, the raw IOPS of drives with different tiers:
- Tier one: 146 GB, 15K FC -- 185 IOPS
- Tier two: 600 GB, 15K SAS -- 175 IOPS
- Tier three: 3000 GB, 7.2K SATA -- 75 IOPS
- SSD: 512 GB card -- 50,000 IOPS
Then, let's consider a 10 TB volume (mostly to make the math easy). To take the most extreme example, let's compare a tier one volume to a tier three volume. In this case, we would need 69 146-GB HDDs to equal 10 TB and four 3-TB SATA HDDs. The tier one volume would yield 10,074 raw IOPS, and the tier three just 300. It just doesn't seem like a fair fight. However, by creating a hybrid tier three-SSD volume with just 5%, or 512 GB, of the 10 TB volume being SSD, we get 50,300 raw IOPS. That's a huge performance leap with just a small sliver of SSD.
To consider cost, let's assume that the 146 GB drives cost $1400 each, the 3 TB drives cost $2900 each and the 512 GB SSD card costs $40,000. This yields a total cost of $96,600 for tier one and $51,600 for the hybrid tier. On a per-gigabyte basis, the price premium for tier one is just under 2X. On an IOPS basis, tier one costs $9.58 and the hybrid tier just $1.03. On top of that, there is far less space, power and cooling for the hybrid tier.
There's plenty of room to quibble about specific drive IOPS, the efficiency of different array vendor designs and so on in the real world. Using 300 GB FC drives would cut the number of tier one drives in half with about the same raw IOPS, but the price per drive would be higher. The comparison would be closer, but still substantially apart on a price-per-IOPS basis. In addition, the prices quoted are intended to be reflective of enterprise arrays, so there's probably plenty of room to argue absolute numbers there as well; the ratios, however, are still relevant. So, the example above is just that -- an example. Like any other formula, plug in the numbers that apply to your situation. The difference is so substantial that the conclusion is likely to be the same; just the justification numbers will change.
A key component to making a hybrid tier work is automated tiering software. This is an additional cost to consider in your calculation. Although random access is not nearly as random as one might think, the vendor's caching-software hit rate is a critical factor in overall performance. Frequent hits to the back-end disk would negate much of the SSD advantage. An application that sequentially reads the contents of entire volume will still benefit from SSD, but not as much as other apps will. Hybrid tiers do require more careful thinking about design, but they make it possible for IT organizations to substantially reduce, if not eliminate, the use of low-capacity, high-cost tier one HDDs.
Automated storage tiering: Higher performance, lower cost?
Best practices for expanding or replacing a disk array
LUN management at the heart of SAN configuration