|
||
|
|
|
              |
|
 |
 |
||||||||||||||
|
 |
|
||||||||||||
By William Van Winkle |
||||||||||||||
Box Choices: JBOD vs. Server As you'd expect, storage connection technologies are only half of the discussion. Next, we have to figure out what boxes to have at each end of the connection. Obviously, there has to be a server, so let's take that as a starting point. Say we have a conventional pedestal system that's been assigned the task of becoming a storage server. A pedestal will often get you six to eight drive bays. That's not a bad RAID setup. But growth happens, and not all pedestals are friendly toward drive hot-swapping, which is something clients want when downtime must be minimized.
At this point, many will opt for a direct-attached JBOD enclosure. A JBOD enclosure shouldn't be confused with a JBOD drive-spanning configuration. You can configure any type of RAID within a JBOD enclosure provided that it's supported by the host controller. Alternatively, some enclosures have RAID hardware built into them. These are called RBODs, and they're less common than JBODs. Not all JBODs are rackmount. AMCC/3ware's Sidecar is a good example of a "starter" JBOD. This four-bay, well-styled shoebox cables out to a 3ware 4-port, PCI Express, SATA II RAID controller (included) that sits in the server. According to AMCC, the throughput for RAID 5 reads and writes can exceed 200 MBps. For those who need little more than to add a small but high-performance DAS RAID in applications such as creative multimedia content, the Sidecar is a perfect fit.
Many DAS deployers, however, will want the ability to scale more. In a regular DAS configuration, a SAS RAID controller card in the server sports an external SAS port. This cables out to the back of the JBOD. JBODs are almost invariably rackmount in form factor, with 2U and 3U being the norm and more affordable price per gigabyte 3.5" drives still being favored over 2.5" models. Naturally, there are some larger JBOD options. AIC/Xtore's XJ 2000, for example, crams 48 drive bays into a 4U SAS enclosure. The JBOD boasts up to 48 Gbps of bandwidth through twin, hot-swappable expander modules and can take up to 24TB of capacity. Keep in mind that a JBOD like the Xtore line is hardware-only. You're still on the hook for coming up with software to manage all of that storage. A conventional SAS HBA will control up to 122 SAS/SATA drives. So even if you were to fill up that XJ 2000 with 48 drives, you could still use the enclosure's backplane ports to connect another JBOD, cascading on more drives until that 122 drive limit was reached. At this point, if still more drives are needed, you'd add another controller into the server for the ability to tack on another 122 drives, and the controller vendor's software should be robust enough to make managing 244 drives no more difficult than the handful inside the pedestal.
Of course, many companies don't need scores of terabytes in storage capacity. For some, 5TB may seem like a bottomless pit for the next three years. Given 500GB drives, that's 10 drives plus maybe two more as hot spares in case of failure. Twelve drives can easily fit in a 2U/3U enclosure. Assuming you want to keep on with the simplicity of a DAS model, you have a few form factor options. You can try to find a tower large enough to fit a dozen drives. You can pick a small pedestal and cable out to a JBOD. Or you can opt for what the industry normally calls a rackmount storage server, meaning a rackmount enclosure with an array of drive bays in the front all interfacing with a midplane board cutting across the chassis. The midplane in turn cables back to a controller mounted on a server motherboard, complete with chipset, memory, display adapter, and everything else needed for a self-contained server system. (We prefer the word midplane to backplane, a common synonym in the industry, because backplane can also mean the rear panel where you find ports and the power supply vent.) The hot ticket in the storage server space for Q2 of this year is Intel's SSR212MC2, commonly known by the code-name McKay Creek. This is a 2U, 12-drive Bensley server sporting the S5000PSLSAS motherboard and support for 32GB of memory, redundant-capable 850W power supplies, and up to two 5300-series Xeon processors, meaning up to eight cores fueling a maximum of 9TB of storage (given the top-end 750GB SATA drives available for validation as of this writing). A 2U, 12-bay server is nothing new, but being able to plant eight processing cores behind them is, and the amount of engineering Intel had to pour into making this possible is substantial. The SSR212MC2, unlike Intel's preceding storage servers, comes with enclosure management software, but no application or OS software. The server features plenty of nifty engineering twists ranging from 10 hot-swappable midplane fans to support for two 2.5" internal boot drives that cable to a motherboard-mounted SAS controller. Bootable Disk On Module (DOM) cartridges can be installed on the IDE, USB, or SATA ports. For $800 and change, you can opt to enhance your SSR212MC2 barebones with Intel's LSI-based SRCSAS144E ("Boiler Bay") SAS RAID controller card.
The SSR212MC2 without Intel's RAID card runs about $2,800, which is very aggressive considering the form factor, processing power, and service Intel provides. Intel's one arguable shortcoming is that it has a very slim storage server catalog. For the second half of 2007, we're unlikely to hear about much from Intel besides this model in the storage server category. Those who want far more variety and perhaps to have their servers arrive pre-built should look to the likes of Supermicro. With hundreds if not thousands of SKUs available, finding just the right part at Supermicro may take more patience and communication, but the company's quality and performance always make the time well spent. Take the 836E2-R800V as a case in point. This is a 3U, 16-drive server chassis, meaning it comes with the case, fans, midplane, slimline DVD-ROM, and 800W redundant power supply but no floppy drive, motherboard, or anything that would plug into one. (Supermicro has validated the 836 chassis for 8-core Bensley configurations.) The move to 3U means that you can now add full-height cards rather than half-height. Supermicro builds in redundant cooling and some other catchy twists, but we like this particular storage enclosure for its integration of two LSI SASX28 expander chips. Each chip supports 28 SAS devices accessible via a backplane port. With no drives or motherboard, the 836E2-R800V rings in at under $1,300, which is a great price considering the two LSI SAS expander chips. You'll still need to come up with a suitable SAS or SATA RAID controller, but at least future expandability, the ability to cascade additional JBODs, and enclosure management (an LED system for assessing status on each drive) all arrive in the box. While vendors and the industry as a whole tend to refer to products like Intel's SSR212MC2 and Supermicro's SC836E2 and the 836E2 as "storage servers," you should be careful not to pigeonhole your products into any given category. The beauty of a modern, multi-core server is that it can be virtualized into several applications. With eight cores at your disposal, you might virtualize three servers on the machine, only one of which is for storage. You could view Intel's platform as a general-purpose server that just happens to have 12 drive bays for data storage. "The JBOD has been one of the things a lot of people have used over the years because a ‘SAN' was just out of the question," says Young. "It was cost-prohibitive. A JBOD was an extension of storage for the server. Remember, three years ago, a big drive was 73GB. Twelve of those enterprise hard drives still wasn't that dense. Now, with 1TB drives, you've got 12TB in something like Intel's new box. You don't need to go external as much now. So McKay Creek can be NAS or a SAN, depending on what you put into it, or just a regular server with a whole bunch of storage in it. Put Windows Storage Server on it and you've got a storage appliance. Put on Open-E or something similar and you've got a full-blown NAS or SAN. Personally, I'm going to put it on my site as a regular server you can put 12 disk drives in." One of the best attributes of the SSR212MC2 and its many competitors is that they're flexible enough to address all three storage models. As stand-alone boxes connecting via SAS, they can be high-performance DAS enclosures. Go through the NIC and you've got NAS. Tack on iSCSI support and you've got SAN. This fact in itself speaks well of the server's ability to meet changing SMB needs over time. The DAS solution of today may need to transition into being tomorrow's iSCSI answer, and the ease with which this can be done may be of the value points behind your storage implementation. Why a server box instead of a JBOD? Good question, and figuring out which side of the fence to stand on will depend in part on the bandwidth required by the user's application. If the customer needs only occasional access to a very large pool of data, then a dumb JBOD enclosure or two is fine. High capacity with low throughput doesn't require much in the way of processing power, only effective management. On the other hand, some storage applications need not only high capacity but also high bandwidth and plenty of processing power. Keep in mind that not only do you have to consider the processing load inherent to an application but then multiply that load across an anticipated number of concurrent users. "There are applications like video-on-demand where you actually want the compute resources in your storage box," says Sonny Banga, storage systems product marketing manager at Intel. "Otherwise, going through a switch to multiple enclosures, your performance is going to degrade. The IP processing carries an overhead load. I was talking with someone who sells storage into movie houses down in L.A. They want their McKay Creek unit to have the compute capability to compress, decompress, compress again, and transmit files. The files they store are compressed MPEGs, and they decompress them before streaming out. Applications like that are where you need compute cycles in the storage box. At some point, it would be cheaper to add JBODs or RBODs onto a server, but you're not going to get anywhere near the same level of performance. You have to go back to the various servers throughout the network every time you need to process a request."
This is the same reasoning behind answering another question: Why buy two 2U servers instead of a single 4U server with twice as many drive bays? Because the compute resources in the box may only be sufficient, given the application, to power the amount and type of storage running in the 2U enclosure. For double the number of drives, the application may need more horsepower than, say, two CPU sockets can provide. Moreover, you have to develop a sense for leaving room for growth in the buyer's storage/compute resources without overbuilding and overcharging him. "Look at what you think your storage capacity growth needs are over time," adds Banga. "If you think they're going up on a hockey stick, you're gonna want to have something with a lot of capacity capability. Even if you're only using three or four disks now, it's nice to be able to scale up to 12. A tower will only generally give you six or eight drives. And look at the 2U form factor. Some businesses have very small server rooms or available floor space, and to be taking up multiple 5U or tower boxes, that can become inefficient over time. If you don't think your storage is growing at a high rate—and I don't know many people who fall in that category, by the way—then maybe you don't need density." If you have 10 or 20 servers all needing access to the data on one JBOD, you have a sad bottleneck in the making. This is where SANs come in. Drop the SAN onto a Fibre or iSCSI network, everybody will have access to it, and the admin gets one central storage server managing all of the data—and that includes granting or denying access to people wanting that data. Of course, opinions vary. Adaptec's Tom Treadway, CTO for block storage and RAID, counters: "A JBOD is always going to be faster. Let's say you've got a Gigabit iSCSI target; that's 100 MBps throughput. On a DAS controller card plugged into a server, the throughput is probably 500 to 800 MBps, depending on what type of card you have. DAS is always going to be faster, but nobody likes DAS because then you have islands of storage. But we still sell plenty of DAS. ISCSI definitely hasn't taken over yet." Choices: What Kind of iSCSI? ISCSI may not have taken over yet, but it's definitely on its way. Or perhaps it would be more accurate to say that iSCSI support is approaching ubiquity. As we've seen, it's no big deal to add iSCSI initiator support to a server. The trick is having iSCSI target implementation in the storage resources, regardless of whether that iSCSI functionality is used immediately. An excellent example for this point is StoreVault's S500, which was designed by the brains from NetApp specifically for the SMB market. The S500 is a 2U, 12-drive box marketed as a NAS/SAN solution, although, from a hardware perspective, it seriously blurs the line between appliance and server. In order to avoid unfavorable comparisons, StoreVault goes to some lengths to mask the fact that the S500 is based on a Celeron processor—and understandably so. With prices for a 1TB (four 250GB drives) configuration starting at $6,000, a side-by-side comparison against a box like Intel's SSR212MC2 outfitted with a pair of 5100 Xeons can't help but look suspect. The thing to keep in mind is that the S500 really is built to be a storage appliance, and because of this it includes many optimizations you're not going to spot in a brief look at bullet points. "The other products out there that count are based on Windows," says Drew Meyer, product marketing manager for StoreVault. "There's a big difference between Windows, which is is an all-purpose operating system, versus Data ONTAP, which is our operating system and is all about storage. Things like the efficiency in the way we use disks to store data—we're dramatically more efficient than a Windows-based product. We can do point in time snapshot copies of data with no performance impact in serious quantity—255 per volume. Windows is roughly 64 per volume, and there's a major performance impact. We can handle two disk drive failures in a single array. No other product can do that without a performance impact." This last point is open to debate. If you refer back to Adaptec's comments at the end of last month's article, you'll see that Adaptec and others have their own ways of coping with RAID 6 processing, and it appears that no one to date has done independent benchmarking of both platforms to see which is most desirable in different user scenarios. StoreVault calls its implementation of RAID 6 "RAID DP" for "Double Protection," a feature StoreVault computes in software through its ONTAP OS. Clearly, even if the claims of zero performance impact versus competing solutions are questionable, having an optimized operating/application system yields obvious benefits. Crunching RAID 6 under Windows with no hardware offload would be crippling.
There are other noteworthy aspects of the S500, such as the 256MB, battery-enabled flash memory card that protects any data caught between the network port and the hard drives for up to three days in case of power failure. There are two Gigabit Ethernet ports, 1GB of DDR2 system memory, dual redundant power supplies, and native RAID 4 support. (RAID 4 is like RAID 5 except that the parity data for all disks resides on a single dedicated parity drive. In general, RAID 4 performs much better on reads than writes.) StoreVault bundles its very accessible, GUI-based StoreVault Manager software package to run from a console system across the network. Those wanting to deploy the S500 as a NAS box get native NFS and CIFS support. Those wanting a SAN solution get iSCSI software initiators for Microsoft, VMware, Red Hat, SuSE, Netware, and Solaris. There are also iSCSI hardware initiators from Silverback and QLogic, and StoreVault offers a QLogic SAN Starter Kit for adding Fibre Channel. StoreVault's objective is to be ready to run however the buyer wants right out of the box. "On top of all that," says StoreVault's Meyer, "NetApp's got a feature built into ONTAP called Block Checksum, where we're constantly monitoring what's written on what sectors on the drives. That lets us know if there's a bad read or if a sector has gone bad. If so, we move the data to another location on the drive. If we get a bunch of blocks that show up on the drive, we can invisibly move the data to a hot spare. The whole point to all this low-level investment is that you never want to throw your customer into a risky situation. Compare that to all the simple RAID 5 arrays out there running off controllers, or even software RAID; you don't get that kind of depth or expertise in the software. When a disk drive starts to go bad, there's no awareness. Suddenly the customer is faced with a surprise rebuild when they walk in the next morning. It's a whole different degree of reliability and security and comfort for the end-user's data." The S500 does have its limits, of course. As per the Intel example above, you wouldn't want to use this box for realtime video streaming. This is not a high bandwidth solution, and StoreVault's software doesn't provide for functions such as virtualizing drive volumes across multiple enclosures, so there are boundaries on the unit's flexibility as a SAN product beyond its bandwidth limitations. Meyer counters that "clustering and failover is not a market we're after" and that "our target audience is not likely to go beyond 6TB." The S500 excels in its niche for low-bandwidth, moderate capacity, widely configurable, easy to install, nearline storage—exactly the sort of thing an SMB reseller would be interested in. As a counterpoint, let's examine something closer to a JBOD, such as D-Link's DSN-3200 xStack Storage Area Network Array. This is a 3U, 15-drive enclosure with eight Gigabit Ethernet ports in the back. (The DSN-3400 swaps these for a single 10 GbE port.) The starting price with no drives is about $6,000. According to D-Link, a comparable unit from the likes of EMC would run $20,000. Like the StoreVault, D-Link's xStack includes secondary cache memory (256MB to 4GB) and a 72-hour battery backup. A hardware RAID controller supports levels 0, 1, 0+1, 5, and JBOD and also allows for RAID level migration. Features such as automatic defragmentation, auto-rebuild of spare drives, and drive roaming (movement of drives to different bays when powered down) are all supported in firmware. One feature available through the GUI is something called the freepool. With the appropriate policy in place, if a drive error is detected, the xStack will automatically start to do a RAID rebuild by pulling a drive out of the freepool to replace the failed drive while also sending an alarm notice to the admin. The xStack doesn't come with a storage operating system, but there is an embedded, IP-based management interface and the enclosure does support Windows versions from 2000 SP3 to Server 2003, with support en route for Linux, Mac OS, and Solaris. From head to tail, this is an iSCSI box. All eight Gig ports can easily be aggregated for a maximum of 850 MBps of bandwidth. D-Link integrates a full iSCSI TOE. Admins can aim multiple initiators at one port or any combination of initiators spread over multiple ports to help control bandwidth. To create a single volume bigger than one xStack can hold, or to span a smaller volume across more than one enclosure, including across disparate locations, you would use the xStack manager plus a virtual machine manager—standard stuff for those already doing VMM implementations. Beyond the Box Managing data as it grows takes SAN-type functionality. We have little faith in any vendor's claim that starts with "all the capacity any SMB will need in the foreseeable future is..." Between RAM's marketing, writing, and design offices, we have well over 10TB of storage up and running, nearly all of it DAS, and the need for us to start bonding these resources into a SAN grows larger every month. We are experiencing a problem now found throughout the entire IT world: storage sprawl. With a handful of storage pools, sprawl may be controllable. But when you get to five or ten servers covering scores to hundreds of employees, the benefits of automated, centralized management with features like hourly snapshots and auto-failover become too great to overlook. These are things that no DAS controller is going to be able to cover adequately. Fortunately, the five- to six-figure enterprise storage solutions of 2003 are now a feasible four figures for SMBs. "iSCSI is supposed to be cheap," says Adaptec's Tom Treadway. "We want to give people a very inexpensive way to build iSCSI-to-SCSI/SAS/SATA boxes. That's why we divorced all of our software from the box, and we're selling what's called a DOM, which is basically a bootable CompactFlash drive that has our iSCSI target software on it. So you can buy anybody's server, plug in one of our RAID cards, and then plug in this DOM, and it boots up the iSCSI target box with management built into it and everything. You've got an iSCSI-to-SAS box with whatever motherboard and chassis you want." If this "make anything into iSCSI" sounds overly simple, that's because, in a way, it is. The only real difference between a NAS device and a SAN device is support for the iSCSI protocol. Factors such as having a TOE in the device can add differentiation between products, but the idea stays the same. Once you've got an initiator and target in place, the rest becomes an issue of management software that will help reign in existing storage and make expansion easy. Bypassing the conventional roadblocks of expansion is what makes iSCSI so appealing, especially to SMBs prone to quick resource scaling. "You set iSCSI up so that it knows where the storage array is and can create volumes, permissions," says D-Link technical marketing manager George Cravens. "All that stuff can then be exported to the LAN. Say you set up a 100GB volume today and in six weeks it's running out of space. You can go in through the administration tools and expand that volume by telling the storage array you want to add another drive into the volume or increase the size across the drives its using. The users don't know anything happened, only that there's more space. It's easier to manage. As far as backups and things, there are companies doing full-on, enterprise-class utilities that'll do snapshots and offline backups. It's totally invisible to the user, but the IT guy can set it up to do periodic snapshots during the day of changes, then every couple days or on the weekend shoot a copy off-site for disaster recovery or whatnot. That level of software package rides on top of whatever your actual storage hierarchy is, whether it's iSCSI or Fibre Channel." Some of this software functionality resides in the target. Lefthand's iSCSI target software, for example, bonds boxes into virtual volumes. So do targets from FalconStor and Open-E. SAN virtualization means that users don't have to know where volumes exist. Everybody might share an M: drive, but nobody would know where it physically exists. Conceivably, the physical disks behind M: might be in three different cities. Marketing can have it's own shared drive. Engineering can have another volume, and so on. Many companies, as they get bigger, do exactly this sort of volume segregation. Fortunately for the channel, there is still enough complexity involved in SMB storage, no matter if it's DAS, NAS, or SAN, that outside experts are often wanted for their expertise in long-term planning and setup. "Smart resellers are telling us some things," notes StoreVault's Drew Meyer. "They're doing migration jobs, which is like step A1, simple and basic. How do I get the data from my 15 file servers onto my new storage effectively and accurately? The reseller can do that. From there, it ranges up to the tape strategy. What about my tape automation and my offsite archive solutions? Resellers can advise on that. Resellers can go further with things like StoreVault replication, which will copy data from one StoreVault to another. You can either set it up in remote offices, or some VARs are even hosting that and providing it as a monthly service for their end-users. So now they've got a revenue stream based on backups happening every night on a scheduled basis. Finally, really sophisticated technology infrastructure overall is where we see people going from an environment that's maybe 30 or 50 people with tower servers scattered around and drives hanging off of various machines. You can replace all that with a StoreVault and a VMWare installation and a Gigabit switch, and they're done for the next five years." The last half-decade has largely been a time for the building blocks of advanced storage solutions to mature and become affordable to the mass of businesses. Now is when resellers are needed to build strong solutions out of these blocks that will help protect clients and allow them to continue growing. There are few segments within the channel that can exceed storage in customizability and high margin. This is an area that grows more accessible by the day and one that you can't afford to ignore. |
||||||||||||||
 Back to top
|
||||||||||||||
Copyright © 2007 RAM Magazine. All rights reserved.
Do not duplicate or redistribute in any form. |
||||||||||||||
