By Dave Simpson
Without fanfare, IBM and Cisco Systems have submitted to the Internet Engineering Task Force a proposed standard for running SCSI over TCP/IP networks. Israeli startup SAN-gate also participated in the specification. If the proposal is ratified, and a number of technology challenges are overcome, it would enable IT organizations to build storage area networks (SANs) based on IP networks such as Gigabit Ethernet.
CLOD BARRERA, Director of storage systems strategy, IBM.
"In the long run, SCSI over TCP/IP could provide an alternative to Fibre Channel for SANs," predicts Clod Barrera, director of storage systems strategy at IBM. However, Barrera and others acknowledge that a number of tough technical hurdles exist. "Until you get the performance questions answered, particularly TCP/IP acceleration, you can't do SANs with SCSI over TCP/IP," Barrera admits.
The performance problems with TCP/IP, particularly for storage-oriented applications, stem from the inefficiency of the protocol stack. "The biggest problem with TCP/IP is the overhead and lack of efficiency in the existing protocol stack," says Richard Lee, president of Data Storage Technologies, a consulting firm in Ridgewood, NJ.
However, Lee acknowledges the potential advantages of running SCSI over TCP/IP: "The compelling value proposition of using Gigabit Ethernet for storage networking is that you already have your core network infrastructure based on Ethernet, so why embrace another technology that doesn't interoperate with Ethernet?"
Others agree that leveraging the installed infrastructure is the most compelling aspect of running SCSI over IP networks. "The main reason for using TCP/IP for storage is that a lot of the existing client infrastructure already uses it for flow control, session management, security, etc.," contends Barrera. "Our view is that TCP does the job, although it's very slow today."
To solve the performance problems inherent in the TCP/IP protocol stack, a number of vendors are reportedly working on hardware accelerators, which would put the protocol stack in silicon on network interface cards (NICs). This would speed performance while reducing overhead and CPU utilization. However, it's unclear what performance level can actually be achieved, and at what price. "The biggest leap of faith you have to make is that vendors are going to make these new protocol stack implementations low cost and widely available," says Lee. "Most users will be non-believers until they see it."
Other analysts also advise exercising caution when contemplating yet-to-be-delivered technologies. "Latency is just as important as raw bandwidth," notes Robert Gray, research director, storage systems, at International Data Corp. (IDC), in Framingham, MA. "You have the raw bandwidth with Gigabit Ethernet, but what about latencies in processing the stack?" Gray notes that the technology is still unproven, but could be very attractive for applications such as backup SANs.
Cisco officials point out that the SCSI-over-TCP/IP proposal will facilitate hardware implementation of the protocol stack. "Our proposal is designed to make it easier to implement TCP/IP in hardware, rather than using some huge processor to accelerate the protocol stack," says Duncan Potter, product line manager for IP switching and storage area networking at Cisco. For example, the proposal specifies methods of encoding information in fixed locations in the packet, "which massively simplifies hardware implementation," says Potter. He adds that NICs with hardware accelerators will probably ship within the next year.
Although IBM's Barrera thinks that Gigabit Ethernet and 10Gb Ethernet may eventually compete with Fibre Channel as SAN interconnects, he predicts that in the near term TCP/IP storage networks will more likely be complementary to Fibre Channel SANs. "For now, think of it as flowing SCSI over TCP/IP up to the edge of the SAN," says Barrera. "The SAN remains Fibre Channel, and you rely on a router."
In this context, according to Barrera, SCSI over TCP/IP would be most useful in providing client access to enterprise storage. "Any client could go through its existing TCP/IP stack, taking SCSI output over TCP/IP instead of a physical SCSI bus."
In addition to the performance issue, another barrier to overcome before SCSI over TCP/IP can be a viable end-to-end solution is that target device manufacturers, such as RAID array vendors, would have to accommodate the new protocol scheme. In the near term, that would require bridges; longer term, it might involve native support, with the protocols implemented in the storage devices.
A number of leading storage subsystem vendors are interested in using Gigabit Ethernet for storage networks, but predict that it will be a few years until the technology is ready for prime time. "SCSI over TCP/IP will be important in three to four years, and we're completely in favor of it," says Jim Rothnie, senior vice president of product management at EMC.
Like Cisco's Potter, Rothnie predicts that protocol stack accelerators, implemented on NICs, will be available within the next year, "so it's time to look at standardization of SCSI over TCP/IP," he says. "Then you'll have the makings of an effective alternative transport for SANs."
Nick Allen, vice president and research director at the Gartner Group consulting firm, in Stamford, CT, predicts that Gigabit Ethernet SANs will be possible in two to three years. However, he adds that Fibre Channel will be the incumbent SAN infrastructure at that point.
"Gartner has for several years said that Gigabit Ethernet will give Fibre Channel a run for its money some time in the future," says Allen. However, his advice to end users is: If you need a SAN today, proceed with Fibre Channel.
Other analysts agree. "My advice is to look at the technology that you can get today," advises IDC's Gray, "and if it can do the job cost effectively, and has the head room for the future, then do it. There will always be a 'better' technology on the horizon."
Regardless of the exact timing of the SCSI-over-TCP/IP standard, and the arrival of hardware accelerators for the protocol stack, it's clear that within a few years at least three underlying technologies will converge on the SAN scene: Fibre Channel, Gigabit/10Gb Ethernet, and InfiniBand (see "Introduction to InfiniBand" in this issue). "What you'll have is a convergence of all these competing technologies in pretty much the same time frame, each with its own strengths and weaknesses," says Data Storage Technologies' Lee.
For more information on the IBM-Cisco SCSI-over-TCP/IP standards proposal, visit http://www.ece.cmu.edu/~ips.
Nishan sketches SoIP SAN plan
Last month, Nishan Systems outlined a bold plan to deliver with its partners a variety of products that will enable SCSI over native Internet Protocol (IP) storage area networks (SANs). Long on promise, yet short on details about its technology and "industry giant" partners, the San Jose, CA-based startup plans to ship products by the end of the year, which may include adapters, access switches, and directors. The technology is dubbed "Storage over IP," and Nishan has trademarked the acronym SoIP.
Storage over IP (SoIP) promises to link a variety of storage/network interconnects over standard Internet Protocol (IP) networks.
Without providing details, Nishan officials claim to have solved the performance problems associated with the TCP/IP protocol stack-without putting the stack in hardware. In addition, they claim that the performance of SoIP on IP networks such as Gigabit Ethernet will be equal to or better than Fibre Channel. Basically, SoIP runs SCSI block-level commands over UDP and IP. Details on SoIP technology are due later this summer.
"We don't see the need to build a second network using a second protocol," says Randy Fardal, vice president of marketing at Nishan, referring to Fibre Channel. However, Nishan says that its approach will preserve the existing infrastructure of Fibre Channel, SCSI, and ESCON devices, and it will require no changes to storage subsystems, operating systems, or applications.
SoIP could also extend SANs to MANs and WANs, providing an alternative to tunneling approaches such as the Gadzoox-Lucent proposal (see cover).
Nishan plans to submit the specification to a standards body within the next couple months. Fardal says that it's possible that SoIP technology could be combined with the proposal submitted to the Internet Engineering Task Force (IETF) by IBM and Cisco Systems.
"The IBM-Cisco proposal is a very early draft submission, and I think there will be more submissions from a lot of other participants," says Fardal. "We're not trying to compete with each other. We just want to get the best of all proposals for storage over IP networks."
Nishan has secured more than $40 million in capital funding, and currently has slightly more than 100 employees. For more information: www.nishansystems.com.