IP SANs: Circling the Data Center? - Part 1.
Server Infrastructure Strategies
Delta 913, 15 May 2001
Sean Derrington
IT organizations can expect a plethora of marketing from non-traditional storage companies promoting new IP-based storage network architectures (transporting SCSI over IP Ethernet networks). Through 2003, users should be extremely cautious about the numerous proposed proprietary solutions, and continue with current storage infrastructure design and implementations.
META Trend: Storage-related costs will constitute 70%-80% of server purchases through 2004+, but as storage hardware prices (e.g., disks, controllers, interconnects) continue to decline 30%+/year ($10-$15/GB by 2003/04), emphasis will shift to storage software, storage-area networks (SANs), and services. By 2002/03, most storage traffic will be offloaded from application networks onto SANs, but interoperability and robust management tools will remain elusive through 2002/03.
Employing storage-area networks (SANs) is becoming a common way for organizations to offload data storage and backup/recovery (B/R) traffic from the application network (see SIS Delta 795, 18 Nov 1999). Fibre Channel (FC) is also rapidly becoming the interconnect organizations are using for server to disk storage and server to other storage device (e.g., tape libraries) connectivity. Additionally, with the dramatic storage growth applications and databases are experiencing, particularly with transactional-based information, organizations' storage budgets (all aspects) are increasing. Moreover, we believe this will continue to be the trend (taking share from server hardware budgets as server costs decline rapidly). By 2004, we believe FC will be the predominant interconnect used to carry storage traffic for 75%+ of organizations' storage/tape library connectivity. However, beginning in 2002 and continuing through 2004/05, IT organizations will have other interconnect options to carry storage and B/R traffic (e.g., SCSI over IP solutions). Through mid-2002, numerous initiatives (e.g., Internet SCSI [iSCSI], Storage over Internet Protocol [SoIP]) and proposed standards (from various vendors) for carrying SCSI traffic over IP-based Ethernet networks will begin to emerge.
Startup companies (e.g., Nishan) and established IP networking companies (e.g., Cisco/IBM, CNT, Adaptec) have begun to design networking solutions that blur the storage interconnect environment (traditionally SCSI-based networks) and the application-based TCP/IP Ethernet networks. The proposed marketing names and underlying technologies of such a network include Cisco/IBM's iSCSI, Nishan's SoIP, CNT's FC tunneling through IP/ATM (UltraNet), and Lucent's FCIP. For the purposes of this and future discussion, SCSI/IP will denote the concept of each approach. At the core of numerous proposals from these companies, which are aiming to garner a share of the enormous storage market (both IT budgets and Wall Street attention), is a new network that carries the SCSI block commands from the server over an IP-based Gigabit Ethernet (GbE) network (in lieu of SCSI block commands being carried over an FC network) to either storage subsystems or tape drives. This design of each proposal will use an "altered" GbE network to initially coexist with, and ultimately attempt to become the ubiquitous replacement to, FC-based SANs. However, such initiatives are not without significant hurdles.
It is also important for IT organizations to note that the new SCSI/IP network will not be the existing Ethernet LAN network. By 2003/04, IT organizations should plan to design a second independent (potentially via subneting) switched SCSI/IP Ethernet network with numerous routing technologies (e.g., FC, parallel SCSI - for tape drive, WAN) for long-term network topology designs, such that application traffic is not impacted.
Network Management
During the past four years, one of the inhibitors to the adoption of FC-based SANs has been the lack of a unified, ubiquitous, end-to-end storage and FC component management (analogous to HP OpenView's Network Node Manager in the Ethernet environment). Despite the historical challenges, Compaq, EMC, Veritas, HP, Brocade, and others have made significant strides in enhancing the management capabilities of FC-based networks, and we believe this will dramatically increase during the next 24 months. One of the primary design goals of SCSI/IP solutions is to leverage the established and mature capabilities of current TCP/IP Ethernet software offerings. Despite the design goal, we believe this will not be a transparent capability for current IP network management software vendors. Instead, we believe there will be customization (e.g., handling storage volume access and zoning) required (the proposals are not using the entrenched TCP/IP specification) to effectively manage the new "blades" that will reside in the "new" Ethernet switches. Additionally, as many of the SCSI/IP initiatives will require routers to convert the protocols for storage and tape connectivity, we believe using current IP network management technologies for an end-to-end view/management will be problematic.
Interoperability
A second design criterion for SCSI/IP solutions is to benefit from the plug-and-play interoperability of existing GbE technologies. However, although that is promised in theory, we believe the practical implementation (and support from storage vendors) will be quite the contrary. First, the hardware components that make up an FC-based SAN (see SIS Delta 757, 15 Jul 1999) will not be the same components as in a SCSI/IP-based SAN. The host bus adapter (HBA) cards that connect the server's PCI bus (and longer-term InfiniBand) to the FC network will be connecting the server's PCI bus to a SCSI/IP Ethernet network. Second, this will often be a separate Ethernet adapter card carrying only storage traffic.
Third, the storage subsystems that currently have undergone significant R&D, testing, and certification for FC connectivity must now be modified (different adapter cards at this level as well) for the new SCSI/IP network interconnect. Within the iSCSI specification, for example, there is a difference between an "initiator" (e.g., a server sending SCSI requests) and "target" (the storage subsystem satisfying SCSI requests). Additionally, we believe the HBA vendors (e.g., Agilent, Emulex, JNI, QLogic) will deliver the target HBAs six to nine months after the first initiator HBAs are completed (early 2002 to OEMs). The other limiting adoption factor for the component vendors will be which SCSI/IP camp they decide to adopt. We do not believe there will be transparency among the many proposals submitted for standardization.
Additionally, the SCSI/IP proposals that have been submitted to the Internet Engineering Task Force for ratification (none of which have received ratification to date) do not support native tape drive connectivity. All tape transports in the market (e.g., DLT, 9840/9940, 3590, Ultrium) either support parallel SCSI or serial FC, but none support native Ethernet. Backup/recovery software vendors (e.g., Veritas, Legato) will have to modify their software to incorporate (and support) the new network interconnect (using routing technologies) while maintaining the highest level of reliability (e.g., media management, error handling, high throughput, overall performance, library connectivity). This will require router companies (e.g., Crossroads, Pathlight/ADIC) to offer another protocol routing option and integration with B/R architectures. Indeed, some B/R vendors still do not support ubiquitous FC connectivity (using FC/SCSI routers) after four or more years of testing. We believe it will be 2004/05 before the component, storage hardware, and software vendors support the different network interconnect to carry storage and B/R traffic.
Business Impact: Storage principles (irrespective of the underlying technologies) are paramount in supporting critical business applications and initiatives.
Bottom Line: Fibre Channel (FC)-based storage-area network implementations, including FC fabric switches/directors, will remain the primary solution well through 2003+ as SCSI/IP standards, interoperability, and integra- ted network management evolve and mature. |
