The SCSI evolution continues: a look at Ultra320 and Ultra640 SCSI. (Connectivity).

The Small Computer System Interface (SCSI) has been setting the standard for storage I/O for over 20 years. From its starting performance of 5MB/sec to its current performance of 320MB/sec, SCSI has continued to evolve through seven generations.

Disk drive performance has been increasing 37% per year. Drives introduced in 2002 have a sustained transfer rate of 68MB/sec. In 2003, they will reach 93MB/sec; in 2004, 127MB/sec; in 2005, 175MB/sec; and in 2006, they are projected to reach 238MB/sec.

SCSI needs a fast system bus. The system bus has to be considered when introducing a new generation of SCSI, because controllers are either two or four channels. Storage data, in most server applications, goes out on the network to PCs and workstations, as well as to memory. This requires the memory bus and the system bus to be four or eight times the speed of the SCSI bus. A four-channel controller can take the full 1GB/sec bandwidth of the PCIX system. When Ultra64O SCSI products enter the market they will have a similar requirement for PCI-X 2, Hypertransport, or PCI Express to handle the data fast enough from the SCSI controllers.

Ultra32O SCSI Paced Data Transfer Mode

Ultra320 SCSI is defined in the ANSI standard SPI-4 (SCSI parallel interface-4). The SPI-4 standard not only doubles the raw speed but addresses the protocol overhead with paced data transfer mode, which includes a free-running clock, driver pre-compensation (pre-comp), optional adaptive active filter receivers (AAF), quick arbitration and selection (QAS), and packetized commands and messaging. Ultra32O SCSI requires packetized commands and messaging, which allows commands and messages to be transferred at the data rate. Several commands can be sent at one time.

QAS allows faster selection of devices, further reducing the protocol overhead.

Ultra32O SCSI drivers are required to have pre-compensation, which drives the first bit after a transition harder by 5075% than it drives subsequent bits. Pre-compensation can be turned off when receivers are used that compensate for the high-frequency roll-off of the cable and backplane. These are the AAF receivers, and they require a training pattern with a low frequency followed by the high frequency to allow the receiver to adjust for the roll-off. The transfer time is 6.25ns, which can be less than the skew of the cable. A third part of the training pattern contains a half-frequency signal to ensure that the receiver compensates for the skew of the system on the right edge. Not all implementations of Ultra32O SCSI require AAF, however, Ultra64O SCSI implementations will require it.

Unlike generations prior to Ultra32O SCSI, the data and clocks transition at the same time and the receiver decides when to clock the data. The training pattern is sent when a transfer series is set up between devices. This continually compensates for changes in the system. The changes can come from devices being added or removed from the system, and from temperature changes. A free-running clock is used with a qualifying signal to provide data-flow control.

Expander communications have been added to allow testing of each bus segment. The SDV (SCSI domain validation) technical report explains how to use the communication with expanders to margin each segment of the bus. Each bus segment can be margined on power up and after a bus reset. The bus is normally only reset on bus errors that should never occur on the SCSI bus. CRC (cyclic redundancy check) errors on a SCSI bus will cause a bus reset. SDV defines how to adjust driver amplitude and the driver pre-compensation level. If AAF receivers are used, the receiver compensation level can only be read back.

For some systems, backplanes and cables had to be redesigned to operate with Ultra32O SCSI. Most systems upgraded from Ultra160 to Ultra32O SCSI must change the controller and disk drives. The frequency of the signals and clocks changed from 40MHz to 80MHz. The high-frequency roll-off was the major problem, as some of the dielectrics caused high-frequency roll-off problems.

Getting a Fix on Ultra640 SCSI

Ultra64O SCSI is under development, and the standard will be ready to support product development in 2003. All the basic features of Ultra32O SCSI will be carried forward in Ultra64O SCSI. Features unique to Ultra64O SCSI:

* Pre-compensation drivers will no longer be required.

* AAF receivers will be required.

* Signals and clocks will double in speed to 160-MHz; the higher speed will lead to a much higher roll-off in the cables and backplanes (the receiver is the best place to compensate for the roll-off).

* The receiver adjusts the clocking for both the rising and falling edges.

* The training pattern will be changed to an alternating pattern between data lines to compensate for some of the crosstalk effects.

* The Ultra64O SCSI total transfer time will be 3.125ns.

* The terminator impedance must match the loading of the backplane, and programmable terminators are used and adjusted in the SDV phase.

When the time comes to upgrade to Ultra64O SCSI, twist-and-flat cables may have to be used in very short lengths. Twist-and-flat cable has high crosstalk, which limits the length of the cables; multi-drop cables will be limited to two meters and three meters for point-to-point. The spacing of the devices can be a problem, and the reflection from the next device can drastically reduce the signal. Devices cannot be placed at quarter wavelengths. The standard 9.85-inch twist and flat cable places devices at quarter wavelength spacing, which creates a periodic structure that has the effect of a notch filter at the key frequency of 160MHz.

Round cables should be used for longer distances and can be used for point-to-point connections up to 20 meters. Better versions of the twist-and-flat cables and the round cables are being developed for longer distances.

Since drives at quarter wavelengths create a notch filter periodic structure at 160MHz, some backplanes will need to have the spacing adjusted. PIP (passive interconnect performance) is a new standard which measures cables, cable assemblies, and backplanes. The SSM-2 (SCSI signal modeling) standard uses the data from the PIP measurements to model the SCSI bus segments to ensure there are no periodic structures that create filters.

Unlike other standards, SCSI allows connection across the generations with expanders to change the different electrical interfaces. A single-ended SCSI-1 device can be connected through an expander to a high-performance LVD SCSI Ultra32O or Ultra64O SCSI bus. SCSI-2 differential devices, also known as HVD (high voltage differential) devices, can be connected through an expander to the high-performance LVD SCSI bus. The LVD SCSI bus can have the full range of Ultra2 to Ultra64O SCSI devices on the same bus segment. Devices operate at the maximum speed that the controller and the specific device will support.

Paul Aloisi is a SCSI Trade Association board member and distinguished member, technical staff at Texas Instruments.

www.scsita.org