Seagate 15K.2 Inflection Point - The New Era of Serial Attached SCSI - Page 4

SCSI Inflection Point: Standardizing on Serial Attached SCSI (SAS) a device's SCSI ID number (fair arbitration modes use additional criteria). Although bus arbitration ensures greater reliability and data integrity, it also degrades performance; the more drives present on the shared bus, the greater the latency for each drive as it waits for the arbitration process to grant it access to the bus. And of course the more arbitration that occurs on the bus, the less the bus is available to actually transmit data, and thus overall performance suffers. Ultra320 SCSI employs packetization and Quick Arbitration and Selection (QAS) to reduce (but not eliminate) the impact of bus arbitration overhead. Limited Scalability In theory, parallel SCSI allows a maximum of 15 devices per 16-bit bus and a maximum of seven devices per 8-bit bus. (In both cases, the SCSI controller itself occupies the remaining SCSI address on the bus.) In practice, parallel SCSI's shared bus architecture results in progressively less available bus access per device as more devices are added to the bus (see Bus Arbitration and Latency, above). As such, the actual number of devices that can be deployed on a SCSI bus while maintaining adequate performance may fall well below 15. Incompatibility With Other Interfaces When the parallel SCSI and parallel ATA standards were written, compatibility was simply not an issue. The roles of the two parallel interfaces were seen in black and white terms, with no expectation they'd ever be called upon to interoperate in common applications. At one end of the spectrum was SCSI, clearly intended for enterprise duty; at the other was ATA, optimized for use in desktop computers. Two decades later, deploying enterprise-class drives (first parallel ATA and now Serial ATA) for near-line and other light-duty storage has become common practice in the enterprise. Parallel SCSI's incompatibility with these popular drive types exacerbates the logistical (and financial) headache of supporting multiple, mutually exclusive interfaces. Cable Expense Early SCSI standards allowed only moderate cable lengths (six meters for SCSI-1, dropping to three meters with several subsequent increases in bus speed). The introduction of LVD signaling (see Point-to-Point Architecture, below) raised the maximum cable length to 12 meters. This came at a price, however, as the sheer number of conductors required (and the substantial shielding necessary to protect them from RFI) contributed to the high cost of premium SCSI cables. Less expensive cables may employ marginal materials and/or construction techniques, with deleterious (and often maddeningly intermittent) effects on bus performance and stability. Internal ribbon cables are highly susceptible to crosstalk (see above), and thus should be kept as short as possible. Bulky Cables/Connectors Parallel SCSI's bulky cabling and connectors clutter enclosures, inhibit airflow/cooling and preclude use with small form factor drives in dense computing environments. Not Hot Pluggable No activity can be present on the bus when SCSI devices are added or removed, effectively dictating that the system be powered down. Obviously, this can have severely negative effects on system uptime and availability. As a result, SCSI drive changes are often deferred to periods of lowest system activity, thus delaying timely deployment of needed resources. Manual Device ID In order for the SCSI controller to recognize and communicate with the SCSI devices on the bus, a unique SCSI ID address must be manually assigned to each device. SCSI ID numbers range from 0 to 7 on an 8-bit bus and 0 to 15 on a 16-bit bus. The address need not correlate with the device's physical position on the bus. Adding drives to a bus without ascertaining the ID numbers already in use can lead to SCSI address conflicts (wherein two devices have the same ID number on the same bus). Such conflicts can manifest themselves as instability and erratic performance, up to and including data corruption. 4