Getting real about iSCSi - Storage Technology Magazine - Page 1
Vendors are starting to give rose-colored iSCSI glasses to their customers. The recent ratification of the iSCSI standard by the Internet Engineering Task Force (IETF) and Microsoft's heartfelt endorsement of the technology is hailed by some as the opening of a vast, new storage frontier. But despite the beautiful portrait that vendors try to paint about iSCSI nirvana, users need to see another picture: What will they experience if they implement iSCSI now or sometime soon?

The view of the iSCSI landscape depends heavily on whose glasses you peer through. Of course, the rose-colored ones cast iSCSI in an appealing light. iSCSI offers organizations the ability to utilize existing networking staff and infrastructure while promising increased storage utilization and cheap disaster recovery solutions.

    Requires Free Membership to View

    Login

    When you register for SearchStorage.com, you’ll also receive targeted emails from my team of award-winning editorial writers. Our goal is to keep you informed on the hottest topics, the latest news and the biggest challenges you face as a storage professional today.

    Rich Castagna, Editorial Director

    By submitting your registration information to SearchStorage.com you agree to receive email communications from TechTarget and TechTarget partners. We encourage you to read our Privacy Policy which contains important disclosures about how we collect and use your registration and other information. If you reside outside of the United States, by submitting this registration information you consent to having your personal data transferred to and processed in the United States. Your use of SearchStorage.com is governed by our Terms of Use. You may contact us at webmaster@TechTarget.com.
The iSCSI performance impact
Does iSCSI slow down server and database operations? The short answer is yes. The TCP/IP protocol--without even considering iSCSI--generates a certain amount of overhead on every server's CPU. Data packets sent back and forth between computers must be built on the host computer, transmitted over the network, received and then broken back down again by the receiving computer. All this activity requires a certain amount of CPU power. However, with the fast processing speed of most systems today, this back-and-forth activity rarely is a problem. Still, the addition of the iSCSI protocol can add a significant amount of overhead to the CPU needed to process the TCP/IP stack.

In addition, Ethernet IP networks that support 1Gb speeds reach their capacity at approximately 70% bandwidth utilization and in some cases as low as 40%. This has to do with the fact that TCP/IP may lose packets during the transmission process. Error checking natively occurs in TCP/IP to ensure the successful transmission of packets occurred. If data packets do get dropped, the sending computer must resend the original packet. This situation only worsens if the network gets congested with too much data traffic.

To mitigate both of these performance issues, two new technologies are emerging. The first is already here. Ethernet Cards that contain TCP/IP offload engines (TOEs) offload the processing of the TCP/IP stack onto the card itself. They use application-specific integrated circuits (ASICs) that are specially designed to recognize the TCP/IP stack and process this protocol before the server's CPU has to deal with it.

The other technology looming large for 2004 comes in the form of 10Gb Ethernet. This helps answer many of the network congestion issues that may arise now. Even assuming a worst case scenario of 40% network congestion in a 10Gb environment, that still translates into network throughput speeds that are six times faster than the best performing 1Gb network today--more than enough to answer almost any of today's or tomorrow's performance issues. Until 10Gb Ethernet arrives, it's probably not wise to consider running an Oracle database with high I/O rates on a 1Gb network.

Yet iSCSI naysayers point to a lack of storage devices--such as tape and storage arrays--that currently support iSCSI. Toss into the mix the poorly deployed storage strategies many organizations have, plus those nagging TCP/IP performance questions and iSCSI can become a dangerous minefield of unforeseen problems.

While sending users into unexplored territory has never stopped vendors from selling products before, anyone seriously evaluating iSCSI must ask--and try to answer--the question of whether iSCSI is a dead-end road? And if your answer is: "No, it's not," the next question to ask is whether now is the right time to head down this path and if not now, when? The iSCSI roadmap answers some of these questions, but doesn't do a good job explaining what storage problems iSCSI creates as organizations try to figure out how to fit this new technology into their current infrastructure.

The iSCSI minefield
iSCSI's value proposition is strong. For example, today many Intel servers natively ship with dual network cards supporting up to 1Gb speeds and plug into Ethernet switch ports with costs running at or below $100 per port. This translates into a cost of as low as $100 per port or $200 per server, assuming a server deployment with redundant paths. Even assuming a worst-case scenario requiring the purchase of two network cards with TCP/IP offload engines for approximately $500 per piece, the cost to deploy an iSCSI Ethernet network still will run only one-seventh of the cost of deploying a comparable Fibre Channel (FC) network--a substantial difference by any measure.

According to Peter Hayden, the CEO of Nashua, NH-based EqualLogic, Inc., the cost to deploy an FC network costs approximately $3,500 per port or about $7,000 to deploy a server with redundant FC connections. While this cost may seem inconsequential for high-end Tier 1 applications, these connection costs exceed the costs of many off-the-shelf Intel servers today.

Of course, iSCSI deployment decisions shouldn't be based on just low hardware costs. Those who proceed without a clear understanding of the management issues associated with storage networking are looking for trouble. Buyer beware: Storage management issues and their associated costs loom large when deploying an iSCSI storage network. EqualLogic's Hayden says, "Hanging the wire helps, but does not solve the fundamental problem of storage networking, which is storage management."

Scott Robinson, the CTO for the Chanhassen, MN-based Datalink says, "iSCSI storage networks will have fairly dramatic different service level expectations for storage traffic than traditional Ethernet networks." He advises companies to dedicate a separate network for storage-related traffic. In addition, the individuals supporting these networks will need to ramp up their storage-specific knowledge.

To manage next-generation iSCSI-based storage networks, organizations need individuals with diverse skill sets. Not only will these individuals need a thorough understanding of current networking disciplines such as routing, quality of service, security and performance management, but an understanding of storage disciplines as well, such as volume and LUN management, virtualization, storage classes and backup and recovery. Storage consultant, Jon William Toigo, says he can count on one hand the number of individuals he knows possessing both of these skill sets.

This dilemma may help explain why so few storage vendors and no tape library vendors currently support the iSCSI standard (see "Who supports iSCSI?," below). However, it's just a matter of time until storage vendors will support iSCSI. Customer demand continues to grow, according to Glenn Clowney, director of strategic marketing for the storage network group at Milpitas, CA-based Adaptec. He reports that many of the major storage vendors have done their internal iSCSI analysis and are starting to deploy the technology in their respective storage devices.

This was first published in May 2003