This is a weird article. It prematurely crowns IP as the network transport of choice for storage networks without making any distinction between core and edge applications, but then it goes on to describe the tremendous technical and standards-related hurdles before IP that practically insure that it will not dislodge FC from the CORE of the storage networks that are being rapidly deployed NOW by the very same companies who already have access to EMC's SRDF over IP which is based on CMNT's SAN over IP. Also, while all this was going on, ESCON, the original SAN mainframe interconnect continues to evolve to the point where it can now extend to 100 kilometers and can now be used to connect between mainframe storage systems and between mainframes and servers. Odd.
A shift in storage
Fibre Channel gives ground to IP for managing distributed data
By Sonia R. Lelii, eWEEK
July 31, 2000 12:00 AM ET
Storage vendors had their chance.
For years, EMC Corp., IBM and others tried to turn Fibre Channel into a de facto standard for SANs (storage area networks). But, unable to agree upon interoperability standards, those suppliers have failed to persuade IT managers to buy in. The lack of a standard, coupled with the increasing need to push data across a WAN, has led some vendors and their customers to explore what they think is a better idea: using the tried-and-true Internet Protocol to manage storage.
"Storage over IP could be a tenfold explosion in terms of how storage can be connected. It will push Fibre Channel into a niche in the data center," said Michael Brown, chairman and CEO of Quantum Corp., a Milpitas, Calif., storage device maker.
SANs first emerged about three years ago as a way to improve a company's access to data by pooling storage devices on a dedicated network, accessible by any server. At the time, observers considered emerging Fibre Channel bus technology to be the perfect transport for SANs because it could accommodate large data packets at far faster speeds, and over greater distances, than the SCSI bus architecture.
But Fibre Channel has fallen short of expectations, a victim of dueling standards efforts that have fractured the promise of interoperability and hampered deployment of SANs. At the same time, the rapid rise of Gigabit Ethernet—whose 1,000M-bps data transfer speeds surpass Fibre Channel's 850M-bps rate—has given IT organizations less incentive to rip out their existing infrastructures to deploy Fibre Channel. In addition, the Internet explosion has increased the need for companies to manage huge amounts of data, not just within an enterprise but across a series of broadly distributed sites.
Enter storage over IP, which proponents say offers time-to-market, management, interoperability, cost and other benefits over Fibre Channel for distributed storage.
"The reality of IP is, it is well-understood by everybody," said Janpieter Scheerder, president of network storage at Sun Microsystems Inc., in Palo Alto, Calif. "Storage needs to become a full member of the network."
WorldCom Inc. is a believer. MCI, before merging with WorldCom, had worked over the past few years to rid itself of the multitude of dedicated links between data centers. Keeping Fibre Channel connections inside the data center, MCI consolidated the external links onto a single IP network — a move that cut support costs for per-megabyte transfers tenfold.
"Today, with the high-speed fiber optic links all over the country, and our entire backbone an IP network, it [made sense to] combine everything over the IP network," said Bob Oliver, the company's chief architect and strategist, in Colorado Springs, Colo. MCI is using EMC's SRDF (Symmetrix Remote Data Facility) over IP.
"[SRDF] was the first step. That allowed us to get rid of the proprietary, special networks," Oliver said. "At one point, we had literally 100 different networks. We had gateways and protocol translators. It was a mess."
Ironically, bringing storage back onto the primary communications network flies in the face of the original SAN concept. The goal was to create a separate storage network to keep high-bandwidth storage tasks from bogging down messaging and other network traffic. "That was the whole reason why SANs were invented, so why are we mucking them up again?" said Shaun Tu, system administrator at the American Security Group, in Atlanta. "The main question is, Why are we doing it? The storage network and the [messaging] network were separated for a reason."
Putting storage on the IP network raises other issues—some technical, some involving standards—that must be addressed before storage and networking vendors make it the connection of choice among storage networks.
Several vendors are proposing to layer SCSI — the predominant storage interconnect in Unix and Windows NT environments — over IP to produce a flexible WAN for storage. Many customers, however, don't consider IP a reliable or efficient enough protocol to handle the robust task of data storage.
IP was designed to handle high-volume traffic in nonsequential data packets that must be reordered at their destination. SCSI, however, needs to have packets in order and has no patience for latency. If packets don't arrive in order, the data is corrupted or the connection fails.
Cisco Systems Inc., of San Jose, Calif., and IBM, of Armonk, N.Y., have teamed up to address that issue. The companies recently delivered a draft specification for the so-called iSCSI technology to the IP Storage working group, an as-yet-unofficial group within the Internet Engineering Task Force standards organization. Last week, Cisco announced plans to acquire Nu Speed Internet Systems Inc., a Maple Grove, Minn., company that's developing iSCSI products.
iSCSI handles the latency problem by placing data in memory out of order. The spec also calls for putting more information in each packet's header so the TCP layer can reassemble packets quickly, thus reducing latency.
This method, however, could pose a problem for certain applications, such as transactional database applications, which, like SCSI, don't tolerate latency.
Another potential stumbling block for using IP to deliver storage data is performance. Processing the IP stack in software alone would put too much demand on a server's CPU cycles.
To solve that problem, startup Nishan Systems is taking a unique approach. The San Jose company plans to ship by year's end an end-to-end storage-over-IP solution that uses IP not just in the WAN but also inside the data center. It will accomplish this by embedding the IP stack into a network adapter, bypassing the server's CPU.
Observers said the longer-term solution is to build IP directly into silicon — a process that could take years. "The basic protocol is the issue, not the transport layer," said Darren Thomas, vice president for the multivendor storage products division at Compaq Computer Corp., in Houston.
Even backers such as MCI understand IP's current limitations for hosting a storage network.
"IP does not have a clue how much traffic is being transported," Oliver said. "You need other tools on top of IP to solve these problems." MCI and EMC are working together to develop policies and directories, two crucial mechanisms for putting data packets back in order after they are transported.
To push their IP agendas, some vendors are partnering to create new technologies for distributed storage. Cisco, for example, is working with Fibre Channel switch manufacturer Brocade Communications Systems Inc., also of San Jose, to encapsulate the Fibre Channel protocol in TCP/IP, enabling their switches to talk to each other. The companies anticipate products in the second quarter of next year.
Separately, San Jose-based switch vendor Gadzoox Networks Inc. has joined hands with Lucent Technologies Inc., of Murray Hill, N.J., to develop a specification, IPFC (IP Fibre Channel), that would route Fibre Channel traffic over an IP network to connect islands of SANs separated by long distances.
Other vendors are looking to take advantage of existing network architectures, such as Ethernet.
Adaptec Inc., of Milpitas, has spent two years developing its new EtherStorage technology and has also submitted to the IETF a draft specification called SEP, or SCSI Encapsulated Protocol, that enables block-based storage traffic to be transferred over existing IP and Ethernet networks.
Regardless of which IP specification wins out, it's becoming apparent that Fibre Channel will no longer be synonymous with distributed storage. Using IP will let companies use a variety of network transports or interconnects for distributed storage.
"Let's remember, Fibre Channel is not the solution. SANs are the solution," said Paul von Stamwitz, principal engineer at Adaptec. "You don't have to be married to a single interconnect."
Looming above the storage-over-IP movement is the standards process. Optimism for a Fibre Channel standard has dissolved as competing groups led by EMC and IBM have rallied supporters for divergent specifications. A similar split could be developing with storage over IP. The IETF's IP Storage working group, which is considering the iSCSI, IPFC and SEP specs, hopes to begin interoperability testing based on a new standard by early next year.
"I assure you we are in a better position than we were with Fibre Channel [when it was first developed]," said Julian Satran, a scientist at IBM's Matan Advanced Technology Center, in Haifa, Israel, and an author of the iSCSI spec. "Obviously, if we go in five different ways, that is not good. But in the end, reason will prevail."
zdnet.com |
