Time To Consider Fibre-To-Fibre RAID Controllers.

During the past year, interest in external RAID controllers packaged in storage enclosures dramatically shifted from SCSI-to-SCSI controllers to Fibre-to-SCSI controllers. SCSI is still the cheapest solution and performance continues to move up. But IT managers are thinking clusters and SANs, and SCSI (at least on the server-side) is a non-starter in cluster and SAN environments due to obvious distance and configuration limitations. During 1999, Fibre Channel became the server-side interconnect of choice for external RAID controllers for VARs selling to Fortune 5000 accounts. Is it time to start thinking about external RAID controllers with fibre on both ends of the controller?

External Storage Arrays

Aside from the very low end of the market, clustering servers with disks mounted in the server enclosures didn't make much sense with two node clusters and will make less sense with four node NT clusters. A single server failure can defeat the entire purpose of clustering the servers-application availability. Some server suppliers are selling clustering software that mirrors data stored on internally mounted disks to disks behind an external controller. If the server fails, at least a copy of the data is accessible by the other server. This seems like an expensive way to solve the shared disk problem in terms of out-of-pocket dollars and server overhead, and will be even more expensive as end-users move to three and four node NT clusters in 2000.

Doubling Storage Needs

NT clustering capabilities will double early next year when Microsoft delivers Win 2000 clusters. End-users will be able to lash together up to four NT servers into a single "system." Even with quad processors, these systems will have over 10GHz of aggregate processing power. Using the "eight to ten drive per node" rule of thumb, these systems will require RAID controllers with shared access to 30 to 40 disk drives. Even with 72GB disks around the corner, "headroom" and the "Internet effect" will drive IT managers to look for RAID solutions on NT clusters that can accommodate 60 to 80 disks. While it might make economic sense to junk disks long before the bearings start creaking and upgrade the entire storage pool to the latest disk capacity point, most IT managers keep the old ones around for a long time and just keep adding new ones as storage requirements grow.

Costs And Fault Tolerance 

When considering the total cost of storage arrays, disks represent the lion's share, assuming storage management costs are excluded from the equation. Fibre disks cost the same as SCSI disks from some disk vendors; other vendors charge a small premium for the fibre version of their product lines. This seems to be the case for 7200RPM and 10K drives. But fibre disks are dual ported and that means a higher degree of fault tolerance and higher data availability.

Consider dual active external RAID controllers with a pair of SCSI buses to the disk pool (Fig 1). Each SCSI bus connects the two controllers to up to 14 disks (one SCSI bus ID is used by each controller). If SCSI bus "A" or "B" fails, all of the disks attached to the failed bus are inaccessible by either controller, and the applications that store their data on those disks fail. Clustering software provides no protection against SCSI bus failures. Aside from disk failures (recoverable by RAID firmware), power supply/fan failures (protected through redundancy) and controller failures (recoverable in dual active configurations), cables are the most failure prone elements in the array and there is no recovery mechanism with SCSI-attached disks. Now, consider the same array configuration with back-end redundant fibre loops (Fig 2).

The disks are dual ported between the redundant fibre loops. If Loop A fails, the attached disks can be accessed over Loop B. In this configuration, the disks are protected (RAID), the controllers are protected (dual active with transparent failover) and the cables are protected (redundant Fibre Channel loops). For mission critical applications, it may be worth paying a few extra bucks for an I/O subsystem with no single points of failure.

Low Cost End-to-End Fibre Controllers

Until recently, the cost of RAID controllers with fibre interfaces to server and to the disks was prohibitive for most NT environments. Two developments have changed that situation and enabled a new breed of external RAID controllers that hit the cost points required on departmental NT servers, clusters, and SANs-declining fibre chip prices and fibre chips with Multiple Target ID (MTID) support.

Lower Chip Prices

The cost of fibre components is coming down. Hardware vendors were paying over $100 for gigabit Fibre Channel chips. Those prices have come down 20% to 30% and will continue dropping as demand builds and two-gigabit chips hit the market.

MTID Support

In the past, fibre RAID controllers required dual fibre host ports to provide transparent controller failover in dual active configurations and each host port required a Fibre Channel chip. One host port on each controller in a dual active pair was configured as "active"; it serviced the I/O requests directed to the controller. The other host port was configured as "reserved" to assume the loop address of its partner's active port if its partner failed (Fig 3).

The current generation of Fibre Channel chips has embedded MTID support that allows a single fibre host port to emulate multiple loop addresses. A fibre host port with MTID support can respond to I/Os directed to its native loop address and, if necessary, its partner's loop address. Hence, fibre RAID controllers can now be designed with a single fibre host port and still provide transparent controller failover. With fewer lower cost fibre chips, it has become feasible to design all fibre RAID controllers that hit the Intel server price points and provide users with no single-point-of-failure disk arrays.

Kevin Smith is the senior director of business management and marketing for external products at Mylex (Boulder, CO).