The Storage Economist

In this second blog on the topic of DAS economics, we can outline one dimension of comparison and that is total cost of acquisition, or TCA. It seems obvious that DAS storage architectures have a lower TCA than SAN-based pooled or enterprise storage. The unit cost of direct attached storage is much lower at the until level than their FC-based equivalent. But there is more than surface-level analysis needed to determine the true TCA in some types of storage architectures. Remember this discussion is on DAS architectures and the comparisons to other FC SAN storage architectures.

Premise 1: DAS disk architecture is cheaper to buy
•    In the SAN costs, overhead can be high with switches, HBA, directors, ISL ports
•    In most SANs you will find modular or enterprise-class Controllers, RAID overhead, Cache, FC ports that add to the $/TB rate
•    Software costs are usually very different in the DAS model since the protection and management functions have moved into the app layer, and are controlled by the host (as opposed to a dedicated storage controller with cache)
•    The unit level DAS disk drive is cheaper, and the basic TCO appears to be lower cost

Premise 2: Price does not equal cost
•    Acquisition cost is only 20% of the TCO
•    HDS has documented/measured 33 different types of ownership costs (download paper here)
•    Let’s not confuse TCA and TCO. There is usually a better TCA storage with DAS but TCO can be very different.

Premise 3: Cloud computing seems to have created a renaissance or return to older DAS architectures
•    Cheaper to build into nodes
•    Removal of RAID and software overhead is popular with some cloud architectures
•    Scaling and data protection is done by adding more CPU nodes (with disk attached)
•    Data management, protection, copies etc. is done at the app level, with local CPU overhead

So what about DAS TCA then? The unit costs are less; the functions have changed with scalability, management, protection overhead moving to a different part of the computer. Can we use the same $/TB metrics or $/CPU node metric as before when comparing other architectures? I have found that when node or cloud based architectures use naked disk or DAS, there needs to be introduced new overhead costs to support the data protection and scalability requirements of the environment. Raw disk is not usable disk unless there are protections and investments to make the TB usable. Take for instance a cloud DAS architecture that relied on many cluster nodes to provide copy protection and drive failure (in place of RAID) protection. These can be very low cost nodes, but the total sum of extra nodes to provide 1 TB of usable capacity may require 2-5X the raw requirement to meet the business needs. All of a sudden you need many low end disk nodes to provide the basic storage function found in a higher-cost counterpart.

In looking at overhead rates for cloud nodes (CPU and disk nodes) at a recent client, we found a cross over point in terms of TCA, when compared to virtual, thin and tiered FC SAN storage architectures. At a low node count (in the hundreds) DAS had a better TCA. There was a cross over point at around 1,000 nodes. Not a very good cross-over, since the data suggests that DAS and SAN achieved relative parity at around 1,000 nodes and that parity was sustained up to about 6,000 nodes.

In the above graphic, we calculated TCA for usable and written-to storage. DAS was a flat-line relative to TCA for usable, but in order to protect data and meet scaling requirements, the written TCA was about 9x the cost. SAN had predictably higher TCA at low quantities, but achieves relative parity at 1000 nodes.

Now this is just TCA, next posting will extend the cost models to include TCO with such factors as power, cooling, management, waste, migration time/effort etc.