All thin provisioning is not created equal

But not all TP implementations are alike. Because of intense vendor marketing that emphasizes each company's "superior" implementation, it's difficult for a user to evaluate and compare products. Ashish Nadkarni, principal consultant at Framingham, MA-based GlassHouse Technologies, says users should consider the following when comparing TP offerings:

2. What's the maximum supported pool size? What seems like a fairly straightforward question was answered differently by all of the vendors contacted for this article. "For LUNs, which have been thin provisioned, the maximum LUN size is 1TB; for flexible volumes it is 16TBs," writes Rich Clifton, NetApp's senior VP and general manager, virtualization and grid infrastructures business units, in an email.
   
   Craig Nunes, VP of marketing at 3PAR, writes that InServ Storage Server (600TB in the S800 model), unlike other thin-provisioning implementations, "doesn't require capacity to be pre-dedicated into separate thin-provisioned pools ... as writes occur to any thin-provisioned volume, capacity is sourced and configured on-demand from the entire free space of the array."
3. 
   
   
4. What's the chunk size? Jerome Wendt, lead analyst and president at Omaha, NE-based DCIG Inc., says the more granular the chunk size, the less overprovisioning will occur. But as small chunks increase, so does the index managing the chunks, which can impact system performance and require more time to manage. Bob Wambach, EMC's senior director, Symmetrix marketing, says "generally, the granularity of chunks allocated from thin pools will not be a significant differentiator in terms of capacity utilization."
   
   Nunes disagrees: "When the unit of capacity consumption is much greater than the size of the write, the efficiencies of thin provisioning are diminished or lost altogether."
5. 
   
   
6. What's the typical oversubscription ratio? In TP, a large amount of storage is set aside for an app, but storage is only released incrementally as additional capacity is needed; that total capacity is said to be "oversubscribed." Rob Commins, director of product marketing at Pillar Data Systems, says "aggressive users oversubscribe their system between six and 12 months ahead of their consumption curve."
   
   "As a rule of thumb," adds DCIG's Wendt, "the typical oversubscription ratio is 3:1 on Windows and Linux servers, and 2:1 on DAS."

5. What about performance? There are differences in how an array's architecture affects the performance of thin-provisioned LUNs, specifically how the system distributes data across the disks in the thinly provisioned pool. A Compellent system monitors performance and, if bottlenecks occur, the array can dynamically redistribute data to less-used disks. Hitachi has published whitepapers about how its thin provisioning can be used to increase performance for IOPS-sensitive apps when a single array or RAID group can't provide the required performance. In addition, some systems can enhance performance by "dedicating cache to specific pools," says EMC's Wambach.
   
   "When a LUN or file system is created using thin provisioning, we contiguously pre-map where the LUN will grow," says Pillar Data Systems' Commins. The firm's Axiom management software, called Pilot, has an expert system that "finds the best place to lay out the next chunk given that LUN's performance priority and app profile," he adds.
6. 
   
   
7. How extensive are capacity management and reporting tools? This is probably the most important question, because you don't want an app to unexpectedly quit due to lack of disk space. Wambach says TP needs to integrate tightly with existing management and business continuity processes, but "not all offerings do this." It should also be simple and quick to add more capacity to a thin-provisioned pool, adjust the amount of host-perceived capacity and automatically add capacity when thresholds are passed. Eric Schott, senior director of product management for the Dell EqualLogic product family, adds that you should be able to convert data between normal and thin provisioning.
   
   Other considerations include whether there can be different thin-provisioning policies on a primary site vs. a mirrored destination site, and across different tiers. And how good is the vendor documentation? NetApp's documentation, for example, has a section on the advantages and disadvantages of different capacity configurations for storing and deleting snapshot copies.