A technology that promises to bridge the worlds of data storage and networking has passed a key point on its path to becoming a usable standard.
The iSCSI specification has been recently finalised by the Storage Networking Industry Association (SNIA), and the Internet Engineering Task Force (IETF) is expected to endorse the standard with minimal changes. According to Steven Bragg, product marketing specialist at StorageTek, the leading vendors were already so close to the standard that companies such as Cisco were already selling compatible products, while others were on the brink of releasing them. "I believe iSCSI solutions are all but here today," he said in mid-September.
EMC product manager Abie Gelbart says "it's still a little premature to get too excited"
iSCSI extends the reach of SCSI storage devices from a few metres to large distances by transporting the commands and data across TCP/IP networks. Affordable gigabit Ethernet components mean the idea of running block-mode disk traffic over TCP/IP LANs is feasible.
The requesting computer encapsulates SCSI commands and data in TCP/IP packets, which are transmitted across the network to the storage controller. The commands and data are extracted and passed to the storage device, which carries out the appropriate operation. Any data returned by the device (eg, the data read from the drive) is encapsulated by the controller and passed back across the network to the requesting computer.
As there is a possibility that these packets will be carried on an insecure network such as the public Internet, the IETF has required that iSCSI include IPsec encryption.
Part of the rationale behind iSCSI is that standardising on a single networking standard— TCP/IP — across the entire infrastructure yields economies of scale from a hardware and management perspective. It also increases the talent pool that can fill a position, as TCP/IP experience is more widespread than other technologies.
But Gelbart warns this could lead to organisational issues for IT management. What structures will be appropriate if the SAN connection mechanism becomes the responsibility of the network group rather than the storage group? There is much more to storage administration than the SAN fabric, but you don't want two centres of TCP/IP expertise. Is it feasible to merge the network and storage groups?
Wayne Glynne, storage business unit executive with IBM, says there should be room for some rationalisation between storage and network groups, but that would depend on the organisation concerned. But it is an issue that should be addressed, because "the most expensive part of a storage network is its management."
How do you know you need it?
The main attraction of iSCSI is that it provides an opportunity to connect servers to a SAN where this was previously ruled out due to the cost of Fibre Channel components. According to Bragg, a Fibre Channel SAN connection typically costs $8000 to $16,000, including the host bus adaptor, a fibre switch port, and storage management software. The reason the range is so broad is that many organisations have a policy that requires multiple paths between each server and the SAN. This sort of budget is regarded as reasonable for mission critical systems, but few organisations are prepared to pay that much to connect every server in the shop.
iSCSI is a much cheaper technology, costing from "a few hundred to a couple of thousand dollars," says Bragg. This makes it a cost-effective way of connecting secondary servers to a SAN. Extending a "double path" policy from Fibre Channel to iSCSI "is a contradiction to what iSCSI is all about—a low-cost infrastructure for secondary servers," he says.
Despite the savings, he says iSCSI is not appropriate for mission critical servers with latency and bandwidth requirements in order to meet agreed service levels. But the cost advantage means IT managers can reap the advantage of storage consolidation across a larger number of servers. iSCSI is "in our view, complementary to a fibre SAN," he says.
According to Bragg, iSCSI is not especially efficient even with gigabit Ethernet, so Fibre Channel will still be attractive for high-performance applications such as those involving Oracle databases.
Gelbart agreed that the initial takeup of iSCSI will be at the low end, as iSCSI over gigabit Ethernet is inferior to Fibre Channel (especially since 2Gbps Fibre Channel is already available). This may change when 10Gbps Ethernet becomes broadly available, he suggests, but 10Gbps Fibre Channel will probably appear around the same time and will again be the more efficient of the two.
Fibre Channel protocols can also be used over IP networks. Fibre Channel over IP (FCIP) allows Fibre Channel commands and data to tunnel through IP networks to connect geographically separated SANs. One disadvantage is that this still requires the use of relatively expensive Fibre Channel hardware.
Another situation where iSCSI's low cost can come into its own is for inter-site disaster recovery links, Bragg suggests. Using Fibre Channel for this purpose can be very costly, and he says that some organisations are looking at using iSCSI solely for this purpose. Gelbart disagrees, saying that while iSCSI is suitable for SANs, FCIP is more appropriate for replication.
In Australia, distance is the greatest challenge to the "storage on demand" model, says Glynne. Fibre Channel cannot reach beyond a metropolitan area, while iSCSI is practically distance-independent and telcos can provide the long-haul IP links.
Glynne observes that FCIP is a less mature technology than iSCSI, and that the performance of the two should be similar. Consequently, organisations should consider the skills they have before making a choice, he says. IBM has a foot in both camps: "the market will decide," says Glynne.
The low cost of iSCSI means pilot projects are affordable, so organisations can consider putting a toe in the water the next time a secondary server requires additional disk space. "We see iSCSI taking a big chunk out of the NAS market," says Bragg.
Although various vendors claimed to have sold iSCSI products in Australia, none could put forward a customer as a reference site for T&B to speak to. This suggests that the projects are either still at the pilot stage, or that the organisations concerned are secretive.
Hardware changes?
Adopting iSCSI currently means installing an iSCSI router to connect an Ethernet LAN to a Fibre Channel SAN, and this approach will likely continue for six to 12 months, Bragg suggests. These routers are typically a fibre switch fitted with one or more Ethernet ports for iSCSI. Bragg stresses the importance of carefully estimating the aggregate bandwidth needed for a project of this kind, as the throughput of such devices varies from around 50Mbps to 120Mbps. "Scoping is the absolute priority," he says. Since iSCSI is such a new standard, it is also important to check that the router and storage vendors have certified the interoperation of the selected products.
HK Tan, marketing manager at Adaptec Asia Pacific, agrees. Interoperability has been "plagued with a lot of problems," he says. Glynne is more optimistic, saying extensive industry experience with IP means it is unlikely there will be many interoperability issues.
The current roadblock in the adoption of iSCSI is the lack of storage units with suitable interfaces, according to Bragg. He says major storage vendors are looking at providing gigabit Ethernet interfaces on disk arrays and tape libraries by 2003. Gelbart concurs, predicting large-scale deployment of the technology by 2003 or 2004 without making any comment about the likely availability of products from EMC.
You get a different story when you talk to IBM. Glynne points out that IBM started selling iSCSI storage last year, including the TotalStorage IP Storage 200i appliance offered in capacities from 108GB to 1.74TB.
Another performance issue concerns the throughput of the network interface cards (NICs) fitted to the servers. It is possible to use free iSCSI drivers with many recent-model NICs, and the performance may be adequate for file and print servers. Application servers may not have enough processor power to spare, in which case an enhanced NIC with a TCP/IP offload engine (TOE) is called for. At present, TOEs are similar in price to Fibre Channel adaptors, but Bragg expects their price will fall. He recommends organisations test the difference in performance between a normal NIC and a TOE in their own environments before making a decision.
Tan says Adaptec's ANA 7711 network accelerator card (yes, that completes the trio of NIC, NAC, TOE) with copper or optical interfaces would be in beta testing by the end of the year. Benchmarks show gigabit Ethernet can consume up to half of a server's CPU cycles, and moving this function into an ASIC on the network card can deliver a significant performance improvement. TCP/IP offload becomes more advantageous as the throughput increases, he says. "It allows you to scale linearly."
Quantum has held back iSCSI support in its storage products pending finalisation of the standard and the availability of TOE adaptors. The company has provided this feature to selected customers for over a year, and expects a general release shortly, according to Mike Sparkes, Quantum SSG APAC product marketing manager. Existing Ethernet storage libraries can be given iSCSI capabilities through software upgrades, he says, but TOE adaptors are needed for performance reasons. Quantum has given a proof of concept demonstration by backing up servers in Israel to storage devices in California, though the company wouldn't recommend this approach for everyday use, he says.
Network Appliance is already shipping a proprietary implementation of iSCSI for its Filer NAS units, according to systems engineer Steve Bracken. "It's just another protocol" for accessing information stored on a Filer, he says. When the standard is finalised, the Filer software will be made compliant and all an owner would need to do is replace the existing server NICs with TOE adaptors and pay the licence fee to activate iSCSI support in the Filer. Data already stored on the Filer will still be accessible, he says.
If all or most of your servers are located in a single room, iSCSI traffic can be carried over an Ethernet that is isolated from your main network. But if the object is to connect dispersed servers to a storage farm, it is likely that SAN traffic will flow over the general purpose LAN. This extra traffic is very bursty, and so its impact on the rest of the network must be considered ahead of time.
Separate network segments for iSCSI are the way to go, according to Tan. He argues that part of the attraction of SANs was to isolate backup and other traffic from the rest of the LAN. iSCSI allows the construction of a IP "cloud" connecting servers and storage devices, "just like the Internet cloud" connecting servers and clients.
Simon Elisha, senior systems engineer at Veritas, says some people argue that given the segmentation of purpose between a LAN and a SAN, you might as well use different pipes for each purpose. That can be done logically rather than physically, but how well that works depends on the exact requirements and the technology being used. Either way, network managers must understand the implications of putting SAN traffic onto the main network.
In Gelbart's opinion, organisations that have already installed Fibre Channel SANs are unlikely to consider converting to iSCSI for at least three or four years, though they may add iSCSI to their SANs to gain the advantage of consolidated storage across a wider range of servers.
But iSCSI opens up a potentially large market for organisations that couldn't previously afford a SAN. Tan sees the emergence of "a mini-SAN market for customers with half a dozen servers." Such organisations typically have no Fibre Channel expertise, but iSCSI uses standard and familiar protocols, and offers proven reliability and ease of network management, he says.
"It'll be a saviour for small- to medium-sized businesses," says Sparkes, while Bracken says "there's a big sweet spot there" for connecting consolidated storage to Windows application servers (such as Exchange Server and SQL Server) that aren't subject to extremely high workloads.
Some people think iSCSI will find a place in big businesses. Many top tier Australian organisations have delayed moving to SAN, says Elisha, and iSCSI provides them with an alternative to Fibre Channel.
Will it be possible to add iSCSI interfaces to existing storage units? The vendors we spoke to wouldn't give a definitive answer, but most of these devices are fitted with modular interfaces to accommodate various network architectures. "Potentially, there will be an iSCSI card," says Gelbart, in which case the storage devices will be connectable directly to an IP SAN rather than via an iSCSI router.
But storage isn't only about hardware. "You need to make sure your software will work with iSCSI," says Elisha. There are plenty of low-level issues that can cause problems, he says, but Veritas has validated its Volume Manager, NetBackup, and SANPoint Control products against the Cisco 5420 router often used to link iSCSI to a Fibre Channel SAN. With the right management software, the administrative impact of introducing iSCSI should be minimal, he suggests. iSCSI is "exciting from a plumbing perspective, but our goal is to make it boring from an application perspective," Elisha says.
At the end of the day, organisations are interested in business continuity and business efficiency, Glynne says. iSCSI can play a part in disaster recovery by enabling long-haul storage connections, as well as presenting an attractive TCO proposition.
"It's a matter of the end user seeing the business benefits," says Sparkes.
| iSCSI Introduction | Products & Considerations |
