Administrators have to make decisions at numerous points in the storage provisioning process. These decisions revolve around such things as the specific RAID level, the size of the LUNs, as well as zoning, masking, striping and concatenating LUNs. For example, if the application requires mirroring for protection, you'll need to provision double the amount of storage and set up the necessary paths.
In general, experts recommend provisioning LUNs at least 30% to 50% larger than the expected size. "This is not wasting capacity," says Jerry Namery, chief technology officer at Winchester Systems Inc., Burlington, MA. "Especially when you're provisioning for database performance, this leaves room to move things around without hitting a performance wall."
Depending on the vendor, you may have more or fewer choices. Let's use LUN sizing, a storage admin's primary decision, as an example. Ideally, administrators want to size the LUN on the basis of performance requirements, efficient capacity utilization or protection. To provide greater performance, for example, you need to provision the LUNs to encompass more spindles.
For example, Forsythe Solutions Group's Arland might recommend provisioning 100 GB LUNs on a midrange array for database performance. But he would stripe that 100 GB across an entire 16 to 24 disk RAID group, which would allow plenty of spindles for performance. "For performance, you want smaller drives and more spindles in your RAID group," he says.
On the other hand, "big LUNs -- anything over 200 GB -- I would reserve for low-performance applications and archival storage and use fewer spindles," adds Arland. In this case, the RAID group might consist of just four to eight drives.
The goal of sizing is twofold: getting the right amount of appropriate storage while "avoiding being stuck with small fragments of storage which kills utilization," says Namery. The trick is to use smaller LUNs, which leave you more flexibility. Fragmented storage happens surprisingly often, usually when someone starts provisioning the array without mapping it on paper first.
LUN sizing also differs from vendor to vendor. "It helps to know which array you're using. Some of the enterprise storage systems allow less freedom of choice when it comes to LUN size than midrange arrays," says Datalink's Anderson.
Similarly, vendors have different limitations on internal striping of the drives. "The high-end Hitachi arrays don't do internal striping," continues Anderson. "You can still do RAID but not RAID-0." If his client needs RAID-0 and the array won't do it, Anderson provisions a smaller LUN and stripes at the host.
The largest enterprise arrays, originally intended for use in mainframe environments, will often come preconfigured with many small LUNs, around 8 GB to 9 GB. "The high-end arrays like EMC's Symmetrix DMX are designed to work best with small 8.4 GB LUNs," says GlassHouse's Weinstein. "It's built into the design and the microcode of the array. The best practice then is to configure the entire array at once into many small LUNs."
These small LUNs can be concatenated into larger LUNs, adds Arland. In contrast, midrange array vendors may sell the array already cut into large LUNs, often 200 GB in size.
With midrange arrays, the best practice is to configure an entire row or shelf of drives at one time, striping the LUNs across multiple spindles, says Weinstein. But don't feel restricted to small LUNs. In the end, "performance has less to do with the size of the LUN than with how many disks it's spread across," he concludes.
When he was at Nielsen Media Research, Stevenson often relied on the prebuilt LUNs that came with the array. "We knew their I/O capabilities," he says. The prebuilt LUNs worked well with Nielsen's standard RAID-5 configuration. "The only question left was how many spindles to put in the RAID set," adds Stevenson.
You'll also need to think about replication early on in the cycle. "When it comes to provisioning for replication, you have two options -- but a lot depends on the manufacturer," says Arland. You can define the replication during the initial provisioning when you'll pair drives for replication. Some midrange arrays, however, allow for more flexibility by letting you specify replication at a later time. "The best practice would be to take care of the replication during the initial provisioning," he adds.
GlassHouse's Weinstein recommends that snapshot and replication be decided even earlier, when you're gathering the initial requirements. "There is only so much capacity available for internal snapshots, so you need to plan for it during the design phase," he says.
Zoning and masking don't require much planning. "Zoning and masking can be done before or after the provisioning or in parallel," says Datalink's Anderson. The only decision is whether to do hard or soft zoning. Hard zoning uses the physical port on the switch. Soft zoning relies on the device's worldwide name, a 16-character string. "The trend today is to use the name. Management tools and switch vendors prefer worldwide name-level zoning," says Anderson.
Carving up the storage
It's only after they have captured the storage service requirements that the pros start carving the physical array into logical chunks of storage. This entails firing up the storage array vendor's element manager, a low-level tool.
GUI-based element managers are easier. "We're new at this. We got a new array earlier this year so we're using the GUI," says Anthony Bergen, manager of server technologies at The North West Co. Inc., a retail distributor based in Winnipeg, Manitoba. The company is using the Navisphere element manager that came with its Dell Inc./EMC Corp. array.
Bergen is using Dell consultants to set up the initial provisioning on The North West Company's new array. The consultants configured the storage into striped meta LUNs, 300 GB in size. Meta LUNs are created by striping across multiple storage volumes to create new, larger volumes. In the case of The North West Company, the meta LUNs are formed out of 50 GB LUNs. "Meta LUNs allow us to grow our storage easily," says Bergen.
Meta LUNs are similar to concatenated LUNs. In both cases, multiple, smaller LUNs are joined to form a large LUN. Meta LUNs, however, are striped across all the spindles of the constituent LUNs, which results in better performance. With concatenated LUNs, data is written to each LUN sequentially, filling the first constituent LUN before moving on to the next, according to GlassHouse's Weinstein. Meta LUNs and concatenated LUNs are particularly useful when the organization wants to expand existing volumes on the fly.
The final provisioning step entails setting the masking to ensure that only the right servers see the newly provisioned LUNs. System admins can then use the volume managers on the servers to find the storage.
Software vendors are touting automated provisioning tools, usually as part of larger storage resource management packages. But many provisioning tools don't work across different vendors' products. "When it comes to automated provisioning, I've never seen it consistently work right," says Weinstein. "The success rate is about 50%. Something always gets left out." The biggest challenge automated provisioning tools face is the sheer diversity of arrays, with each having to be addressed natively at a low level.
Automated provisioning, it turns out, isn't very automated at all. "The tools essentially use workflow to automate pieces of the manual provisioning process, but some things are always missing, such as change management, approvals or the server side," says Robert Passmore, vice president, research at Stamford, Conn.-based research firm Gartner Inc. They also don't get deep into the storage. "The tools operate at a high level, after the RAID sets have already been created, for example," he continues. "They can't do all the nitty-gritty work. For that you need each array vendor's tool. But, eventually, automated tools will get there."
Automated heterogeneous storage provisioning tools from vendors like CA Inc., Hewlett-Packard Co., IBM and others will eventually make storage provisioning a snap -- but not today. Until then, storage administrators will just have to provision storage the old-fashioned way ... by carving raw disk capacity into LUNs by hand.
Click here to return to The right way to provision storage, page 1.This article first appeared in Storage magazine's October 2006 issue.