ISCSI gains strength from servers

More powerful servers are helping ease some performance concerns for users considering IP SANs, but others say they still want to see more improvement.

In the early days of iSCSI adoption, the fact that IP storage could use existing network and server infrastructure was something of a double-edged sword. Users often saw the benefit of standardizing an entire IT department on the IP protocol in terms of simplicity, but servers were simply not up to the task, necessitating another piece to the puzzle -- TCP/IP offload engine (TOE).

But according to Enterprise Strategy Group (ESG) analyst Tony Asaro, TOE is a thing of the past, thanks to leaps in server processing power. "As CPU and memory get faster on servers, there's less of an impact from processing iSCSI," Asaro said. Also boosting iSCSI's integration into the traditional IP network is the fact that Microsoft's iSCSI initiators now natively support multipathing and remote networking -- two things that were formerly done on TOEs.

"HP's [Hewlett-Packard Co.] DL server line offers offload options for integrated gigabit NICs [network interface cards]," said Luke Kannel, senior Windows server specialist, information systems for a large healthcare company in the Midwest, who asked that his company not be named. "The server CPUs handle the actual payload, but our current CPU utilization on many servers is still under 20%."

Kannel said his company has been using HP EVA arrays with Cisco's 14/2 IP services module in an MDS 9509 storage area network (SAN) switch to present EVA5000, EVA8000 and ESL286e tape library to hosts over IP, most recently in order to do iSCSI boot with PCs. "It allows our [workstation] PCs to be deployed, diskless, very rapidly and failed critical PCs to be recovered within a matter of minutes," Kannel said. "We're trying to leverage the features of our SAN storage arrays with the cost-effectiveness of IP storage."

Other users said they can see a use for IP storage when it comes to replicating for disaster recovery (DR) or backup over long distances, since it's less sensitive to distance than Fibre Channel (FC) or memory protocols like InfiniBand.

Debate continues over performance

ISCSI and IP SANs have been around long enough to gain some credibility -- but they've also been around long enough to see cases like that of OfficeWare Inc., a business equipment dealership based in Kentucky.

According to OfficeWare's chief technology officer (CTO), Chris Resch, the company deployed an IP SAN last year, choosing it because of its lower cost compared to an FC SAN. Resch said his company soon regretted the decision.

"We were trying to back up large VMware Inc. system files, 50 and 80 GB," Resch said. Trying to send those files over IP, he said, ballooned backup windows from eight hours to as many as 18 hours in some cases. Resch said that the Microsoft IP stack (OfficeWare is a completely Windows shop) has limits on the number of IP packets that can be transmitted at a time. But, he said, the problem wasn't solely Microsoft's, and it wasn't because of the VMware application.

"Any large file streaming over Gigabit Ethernet (GigE) is going to give you problems," Resch said.

According to Asaro's presentation "The State of iSCSI," at the recent Storage Decisions conference in New York, just 17% of some 511 companies surveyed by ESG already have an IP SAN deployed, and just 20% of the respondents plan on implementing one soon.

Furthermore, according to the ESG numbers, 59% of the planned adopters of iSCSI and 24% of those who say they have no plans to deploy IP SANs cite "performance concerns" as the reason. Roughly 26% of the "nonadopters" said they would only consider iSCSI if there were documented improvements in the performance of the protocol.

"I'm sure you can architect a solution in either [the FC or iSCSI] realm that would suit your needs," said Tom Becchetti, senior infrastructure engineer, information technology, who also asked that his company, a national financial institution, not be named. "The question is, how easy or complicated is it going to be?"

Becchetti, who currently uses an EMC Symmetrix array and mainframes for high-volume processing, said he's been "playing with" iSCSI using EMC's Celerra network attached storage (NAS) box. What he's liked about it so far is that IP incorporates more virtualization into the protocol -- especially device address virtualization.

"With IP, you're not as bound to a physical device as you are with a channeling protocol like FC," Becchetti said. "I can definitely see a future implementation for iSCSI storage in our environment that uses a farm of different machines with virtualized addresses, so a server could 'check out' an iSCSI tape drive and then 'return' it when it's done."

Hurry up and wait for the next tech refresh

Other protocol enhancements, such as 10 GigE and InfiniBand, hold the promise of solving iSCSI's performance issues.

There's not much of a reason why iSCSI storage can't function in high-performance environments today with the addition of InfiniBand, according to Al Davis, systems manager with MIT's BioImaging center. The lab, which collects, studies and stores high-resolution microscopic images of cells, is phasing out an old Silicon Graphics Inc. (SGI) FC RAID system in favor of "hybrid" iSCSI-over-InfiniBand storage.

"InfiniBand and iSCSI are not nearly as mature as Fibre Channel," Davis admitted. "But we don't always need 24/7 guaranteed uptime for long-term storage. What we need is huge throughput. With 20 Gbps the promise is there for better performance going forward."

In support of applications like Davis', the InfiniBand Trade Association last week ratified the iSER specification, which will allow iSCSI systems to natively connect to InfiniBand. But other users said they're still waiting for 10 GigE instead, because as a memory protocol, InfiniBand lacks the flexibility over distance that is a key benefit of IP.

However, once 10 GigE arrives on the scene, it'll mean another leapfrog over what server CPUs can handle, meaning the return of TOE, according to Kannel. "In our environment, I don't think TOE will be implemented until 10 GigE is deployed to the servers, and I don't see that happening for at least a year," he said.

Still, the greatest wave of adoption for iSCSI may be yet to come, for the simple reason that users are waiting for older FC equipment to reach the end of its life or depreciate before considering IP storage. Some 74% of nonadopters in ESG's study identified "existing SAN/NAS meets our needs" as the chief reason they were not considering iSCSI at this time, and 49% said they would consider it once they see full depreciation on current storage assets.

By then, Asaro said, if the past year is any indication, the protocol will be mature enough to go up against FC in all but the highest level environments.

According to Asaro, a self-fulfilling cycle has begun to take effect in which more implementations of iSCSI (ESG counts over 20,000 so far) mean bugs get ironed out and IP storage gets sharper, which drives up adoption further.

"In the high end of the market for Tier-1, mission-critical storage, they're going to continue to use Fibre Channel," Asaro said. "Eventually, everyone else will be up for grabs."

Read more on SAN, NAS, solid state, RAID