|
||
|
|
|
              |
|
 |
 |
|||||||||||
|
 |
|
|||||||||
By William Van Winkle |
|||||||||||
Switch to SAS? Seriously?
Despite taking a distant #2 in the hard drive market behind the Maxtor-enlarged Seagate, Western Digital is a long-standing, respected name in the business, and we continue to be bemused by the omission of SAS from WD's offerings. Hitachi GST, descended from master enterprise vendor IBM, only offers one SAS model: the Ultrastar 15K147. Adapter houses like 3ware and Promise don't have a single SAS card in their catalogs. Do these companies know something we don't? "We're watching the SAS market," says Western Digital's Steve Shattuck. "The SCSI and Fibre Channel markets have not been a growing space, just pretty steady. In 2000, WD shed a lot of its businesses that weren't as profitable and focused just on desktop. From there, we grew and have branched out into mobile drives and SATA enterprise. But the SAS market is changing, so right now we're just keeping an eye on it." Shattuck also notes that five or six years ago, enterprise storage was dominated by SCSI and Fibre Channel drives. Now, about 30% to 35% of the enterprise market is using SATA. Add to this that we've heard conflicting comments about SAS form factors. Shattuck states that SAS is primarily moving to the 2.5" form factor owing to the potential performance benefit when you address more spindles per rack. Two or three 2.5" drives can sit in the same amount of space as one 3.5" drive. So when you need to fetch a piece of striped data, you've got more spindles to turn and retrieve it. Why stripe to only four drives in four 3.5" bays when you can strip to eight or twelve? Well, one reason may be cost. Another two might be thermals and vibration. The drives' environment had better be pretty solid, and the buyer must need some extremely fast performance and/or local redundancy. Other industry insiders aren't so sure that 2.5" SAS will thrive beyond top-end niches. Sonny Banga, storage systems product marketing manager at Intel, feels that pricing will impede 2.5" SAS adoption for at least the next two to three years. From his seat, he still sees 3.5" SAS having clear dominance. What may clinch the success of 2.5" drives for business storage is the emergence of 2.5" nearline SATA, just as we've seen in the 3.5" format. Today, 2.5" nearline doesn't exist. You have "notebook" drives (Seagate Momentus, WD Scorpio, Hitachi Travelstar, Fujitsu MH series, etc.) and a scattering of SAS models from Seagate's Savvio and Fujitsu's MA families. So far, in the 2.5" form factor, there's nothing in the middle. In the ever-popular scenarios of good, better, and best, the mainstream enterprise server market is still waiting for something "better." We suspect you won't be waiting for long; there's just too much growing need in the market. "Somewhere around the middle of this year, we should see SAS taking over the larger share of the interface market from SATA, parallel SCSI, and Fibre Channel in the enterprise space," says Seagate's Whittington. "If it's Fibre Channel, SAS, and SCSI, at the moment, they're sharing the market at about 30% each. By mid-year, SAS will be maybe 35 or 36 percent." That may not sound like a landslide trend, but it's indicative of where the market is leaning. In the context of storage, SAS looks unlikely ever to be your volume leader. We wouldn't be surprised to see a business world ratio emerge of 10 SATA drives selling for every one SAS. But SAS is increasingly likely to be the crowning layer at the top of your tiered storage solutions, the zone of most critical performance, and as such it needs to be at the front of your sales and deployment strategies.
Serial Controllers Especially in the world of servers, CPUs are meant to enable applications, not storage. Yes, you'll find many options among server boards with integrated RAID, including for SAS, but when push comes to shove, smart integrators will let the CPUs do their job and instead offload storage crunching onto dedicated storage controllers. Note that "RAID controllers" have onboard silicon designed for handling most of the RAID computation. It's not uncommon for a RAID card to knock CPU utilization down from 30% or more to under 5 percent. In contrast, host bus adapters (HBAs) are also add-in cards, but they rely on software and the CPU to perform their RAID tasks. In some circles, this gives HBAs a dirty reputation, but in low- to mid-level environments, host bus adapters can save money versus true RAID controllers, deliver a similar feature set, and usually offer a far better solution than most motherboard-integrated RAID solutions. The case for HBAs would seem to get only more attractive as CPUs gain ever more computing horsepower. You might even question if a RAID controller would be needed if the customer already had, say, a quad-core Xeon or two in play. But again, we come back to future-proofing. As CPUs deliver more power, applications inevitably expand to make use of it. Dual- and especially quad-core chips tend to land in high-load settings. Throw one or two such apps at those processors, particularly when virtualized servers are involved, and all that unused CPU capacity just waiting around for RAID computation suddenly vanishes. Alternatively, you can argue that having a robust storage processor reduces the need for higher-end CPUs. In this light, a RAID controller may pay for itself just in offsetting high-end system processor premiums. Once you get past the idea of CPU offloading, storage card discussions quickly shift to feature support and software-based management features. We're going to leave the management discussion for next month, but one of the key features that helped propel the corporate storage card market was port counts. At a time when most motherboards would top out at four attached drives, perhaps six or eight if you didn't mind them not all coexisting in the same array, high port count SATA/PATA cards had tremendous appeal for those doing security, NAS, high-performance computing, video editing, and similar applications benefitting from sizable arrays. The interesting part about these cards is that they spawned a new type of system. These weren't general purpose servers; they were single-purpose appliances, often counted by analysts as external storage devices. The handy thing is that the hardware in such appliances could stay constant, while the software, and specifically the storage management software, could skew the box toward this or that market niche. Storage card vendors and ISVs figured this out, and the smart controller manufacturers started designing proprietary hardware elements to facilitate these specializations.
"We built a lot of our features and hardware IP around the I/O environment in vertical applications," says Scott Cleland, director of marketing for AMCC (formerly 3ware). "What differentiates us is proprietary hardware built around very high, sequential transfer rates and very large capacities of SATA drives. The other guys—Adaptec, LSI, promise, etc.—take commercial, off-the-shelf I/O processors and differentiate with firmware and software only. We differentiate with hardware, firmware, and software, and that's where the name 3ware came from. So we build a lot of high-speed data movement inside our chip, which the other guys do in firmware, and that gives us performance advantages. We have advanced caching algorithms that are geared directly towards multiple streams. So if you could visualize security cameras around a casino, all streaming full color, motion video into an appliance, we build the firmware and caching features around the fact that each one of those streams is going to require a guaranteed bandwidth. We've got specific features that dynamically allocate cache to those streams so that overall system throughput is faster. We do the same thing with many other types of applications. Because, like with video, if you don't have that, you could drop frames, and that's a big quality-of-service problem. So we've done a lot of work around predictable read latencies in SATA environments." Controller cards offer another key advantage over software-based RAID: system-to-system mobility, sometimes called RAID roaming. When you create a software RAID, you are dependent on that system's operating system being up in order to access your data. If the operating system fails, there's no way to recover that data unless you rebuild the OS and reconfigure the storage in exactly the same way. With hardware RAID, you can move the array anywhere. Simply plug it into a controller on another system and the array will come back up. We're all familiar with corrupted OS installations. RAID roaming is one more value point and level of insurance you can offer buyers that will help improve storage integrity while boosting the system's value.
Not least of all, hardware-based RAID controllers tend to be more expandable, especially SAS controllers. Once you hit the drive count limits of onboard RAID, you're done, even with a paritcularly flexible integrated RAID scheme like NVIDIA's NVRAID. Alternatively, take a mid-point unit like Adaptec's RAID 3805 ($575). Not only does this low-profile PCIe controller card support more RAID levels than you can shake a stick at, it delivers eight SAS/SATA ports out of the box and, through SAS expanders, can support up to 128. Of course, beyond a certain point it will make sense to bring in another controller card rather than overburden one controller with too much RAID computation, but in theory one card should cover just about any SAS/SATA array you can envision. This will entail adding at least one more volume to the storage configuration, but in most circumstances, pooling together (or "virtualizing") large arrays into a single uber-volume is frowned upon by admins. Adaptec adds more value through features like specialized RAID 1E and 5EE modes, copyback hot spare support, staggered drive spin-up, redundant path failover, and RAID level expansion (including RAID 6) with no rebuild downtime. As we'll see next month, this is a little-known benefit that offers excellent price/performance value versus higher-end, pre-configured storage systems from name brand vendors. SAS expanders are one of your greatest storage assets, and they work equally well for SATA drives. With expanders, you can fan out to dozens of SATA drives, yielding a low-cost storage array without a major penalty in controller hardware. Expanders also grow the possibility of being able to maintain separate SAS and SATA RAIDs off of a single controller. (You could even mix formats in the same array, but because an array will drop to the level of its lowest-performing member, why would you toss out the value of SAS units by introducing a SATA disk to the RAID?) This is the beautiful, dual-format, tiered storage promise of SAS. The reality, though, is that there are still some interoperability issues left to sort out. For anyone who remembers the rollout of Gigabit Fibre Channel, this will be old hat. The specs for SATA and SAS are rigid in some ways and open to interpretation in others. These little differences can result in slight signal timing differences that cumulatively add up to trouble as you try to coordinate backplanes, chassis, expanders, RAID controllers, and drives all together—with two different drive types regulated by two different industry bodies. In the real world, we don't expect to see global compatibility between SAS and SATA infrastructures until at least 2008. We mentioned above that RAID 6 was a key method for dodging double-failure risks and that motherboard chipsets to date don't offer RAID 6 natively. This feature is another hot selling point for discrete storage cards, but be aware that all RAID 6 approaches are not implemented equally. "RAID 6 was a big buzzword for the last 18 months," says AMCC's Cleland, "and now that everybody's got it, hardly anybody's implementing it. But here's the thing about RAID 6: There's a penalty to be paid of up to 30% from a RAID 5 write to a RAID 6 write due to the fact you have to do multiple disc operations and multiple memory interrupts in order to get your parity calculations all done. So we looked at that problem and solved it in hardware by doing simultaneous parity calculations. This reduces the amount of disk operations almost in half, and this serves to make RAID 6 more transparent. You can implement a higher level of data protection without suffering any data performance hit. Our 9650 controller, shipping today, and all future controllers will have that in them." The issue of resource utilization is a hot one in many camps, for businesses and vendors alike. While examining different storage alternatives, there were times when vendors would claim superior performance in this or that scenario. A vendor might maintain that its solution offers peerlessly low impact on system resources. Such statements always make us a bit dubious, and since you see "overall system performance" cited so often in storage circles, we took the question to Tom Treadway, Adaptec's CTO for block storage and RAID. "The short answer," he says, "is that we typically see 1% to 5% utilization for all RAID levels, including RAID 6. The reason is that our driver doesn't have to do any of the heavy lifting of RAID processing; it isn't even aware of which RAID level is being used. All the work is offloaded to the controller. The long answer is that it's very dependent on access pattern and controller speed. It's intuitive backwards, but the higher the performance of the controller, the higher the utilization of the x86 CPU. For example, there are cases where our CPU utilization is twice as high as the competition, but with that small slice of CPU time we're actually moving three to four times as much data. So while our CPU utilization might be higher, our CPU efficiency is much better. The reason is that with slow controllers the OS spends more time waiting for an I/O to complete. The amount of time spent in the driver is a fixed amount. Therefore if the controller dropped to near zero performance, the CPU utilization, or the percent of time spent in the driver, would also drop to zero. So in an access pattern that has low performance, like a random access test to a limited number of drives where the controller is basically just waiting on drives to seek and rotate, the CPU utilization will be very low. But in a short sequential read test where the I/O is coming out of a controller or drive cache, the CPU utilization will be relatively high. In fact, it could be much higher than the 5% I quoted above, but it's somewhat of a silly test because CPU utilization is somewhat meaningless." Once again, part of your value as a reseller is being able to educate buyers on what the real world benefits of a solution are and whether this or that deployment offers maximum bang for the buck. Just because a vendor has a known name and lofty claims doesn't mean that its value will hold up under close scrutiny. Building Toward a Solution So now you're hip to the respective strengths of SATA and SAS and know that strong controllers are essential to deploying these drives. These are the preliminary building blocks of any modern storage solution, your bricks and mortar. Now cast a fresh eye on your clients and consider how their current storage needs might be improved through tiering, expansion, and consolidation. Once you've mastered these storage building blocks, your next job is to craft them into customized, scalable solutions focused on your clients' unique and expanding needs. |
|||||||||||
 Back to top
|
|||||||||||
Copyright © 2007 RAM Magazine. All rights reserved.
Do not duplicate or redistribute in any form. |
|||||||||||
