PACKETIZED SCSI: The Need For Speed.
"SCSI has long been the cornerstone of the server hard disk drive business even as new interfaces have been introduced," said Dr. M.S. Bhat, director of server HDD and business line manager. "Ongoing enhancements have allowed SCSI to grow, mature, and to continue to be an excellent interface choice."
Further driving the SCSI performance evolution is the exponential increase in data storage requirements for servers, in part resulting from the widespread use of the Internet and corporate databases that often incorporate multiple data types. In fact, SCSI is now being used as an interface in Storage Area Networks (SANs), in clustered servers, or workstations, as well as for high-speed host-to-host connections. SCSI has also become a preferred interface for data-intensive functions, like video editing and transaction processing. For today's users who depend on increasing numbers of servers and workstations, faster transfer rates are critical. So if you feel the need--the need for speed--look to the next generation of the small computer system interface, packetized SCSI.
WHAT IS PACKETIZED SCSI?
Packetization is an optional feature in Ultra3 and Ultra160 SCSI and will be a standard feature in future devices based on Ultra320 SCSI. Packetized SCSI allows commands, messages, status, and data to be transferred between SCSI devices, like peripherals and storage, faster and more efficiently. Like the traditional parallel SCSI, packetized SCSI is a transfer medium, moving information and data without affecting content or meaning. However, unlike its traditional counterpart where different phases are required to communicate different types of information between devices, packetized SCSI communicates using only two data phases, one in each direction. Although packetized SCSI is a radical departure from traditional parallel SCSI, the two transfer protocols are typically compatible. In fact, a single SCSI device could talk to another device using packets during one operation and seamlessly switch to parallel transfer to perform an operation with a third device.
Packetized SCSI offers new benefits that support user needs, which are being driven by the rapid deployment of servers and workstations with faster data transfer rates. The industry is meeting users' needs with products that provide performance, as well as reliability, interoperability, and backward compatibility with today's SCSI devices. At the client level, the huge installed base of SCSI devices and its continued popularity means we need to add more usability and performance for desktop systems. Packetized SCSI can be a key part of the equation.
WHY USE PACKETIZED SCSI?
With traditional parallel SCSI, all protocol timings are fixed and unable to increase as data rates increase. Thus, even though SCSI data rates have increased from the original 5MB/sec to 160MB/sec (Ultra3 SCSI), the protocol overhead has not changed. With the advent of data transfer rates at 320MB/sec and faster, too much of the SCSI bus bandwidth is used up with protocol overheads. The solution is packetized SCSI. There are a number of reasons why packetized SCSI is superior to traditional parallel SCSI.
PROTOCOL INFORMATION, DATA TRANSMITTED AT NEGOTIATED SPEEDS
With traditional parallel SCSI, protocol information--commands, messages, and status--is moved using a narrow 8-bit bus at asynchronous rates, the slowest method of transferring information. In contrast, data is transferred at the fastest synchronous rates and bus widths negotiated between any two SCSI devices. Thus, the time needed to transfer protocol information remains constant, becoming an increasingly larger percentage of total command time. Packetized SCSI, however, can transfer commands, messages, and status at the fastest possible negotiated data rate. As a result, the time needed to transfer protocol information decreases as the synchronous data rate increases.
MULTIPLE COMMANDS IN SINGLE CONNECTION
With traditional parallel SCSI, each time SCSI devices communicate, only one command at a time can be issued. To issue subsequent commands, the two devices must break the existing connection and establish a new one, often waiting until the SCSI bus is free. With packetized SCSI, one device can send any number of commands to another, significantly reducing the time required to issue multiple commands.
DATA, STATUS FROM MULTIPLE COMMANDS IN SINGLE CONNECTION
Traditional parallel SCSI doesn't allow both data and status for a single command to be sent during a single connection. It can't transfer data or status for a new command without disconnecting from the current command, waiting until the bus is free, winning control of the bus, and sending the proper messages. With packetized SCSI, a device can send or receive data, as well as send status for any number of commands, allowing the most efficient use of the SCSI bus available bandwidth.
When a command fails with traditional SCSI, the device originating the command must issue another command to get information about the failure, delaying any error recovery action. In contrast, packetized SCSI returns all available failure information as part of the status at the end of a failed command. No additional command must be issued and any error recovery actions can take place immediately. In addition, with packetized SCSI, Cyclic Redundancy Check (CRC) is automatic, protecting all transmissions.
HOW DOES PACKETIZED SCSI WORK?
With packetized SCSI, four information units replace all current SCSI phases:
* LQ is the information unit that precedes other information units to indicate the type and length of the information unit to follow.
* Data is the unit that actually transfers the data for an I/O process from initiator to target. The target device can indicate one data information unit at a time and/or multiple write information units to follow.
* Cmd is the information unit that transfers the command descriptor block and the task attribute or task management function from initiator to targer. The initiator device can indicate another cmd information unit coming or a final cmd information unit.
* Status is the unit that transfers the status and auto-sense information for an I/O process.
THE PRESENT: PERFORMANCE IMPROVEMENT
Performance improvement with packetized SCSI can be significant. The following two examples assume that all SCSI devices always have data ready to transfer with no latency delays.
The first example (Table 1) shows the amount of throughput that can be achieved using packetized SCSI over parallel SCSI. Ten commands, four write and six read, are transferred and executed with the SCSI bus running at various rates, from 20 to 320MB/sec. Three different data transfer sizes (2, 8, and 64K) are used for each rate.
* The bottom part of the bar represents the maximum throughput with traditional SCSI.
* The middle part of the bar represents the maximum throughput with packetized SCSI.
* The top part of the bar represents the maximum throughput with no protocol overhead.
* As Table 1 shows, the relative improvement between traditional and packetized SCSI is significant at data rates above 80MB/sec.
* The second example (Table 2) compares five different workloads on a SCSI bus running at 160MB/sec:
* 6 reads and 4 writes to a SCSI device with a queue depth greater than 10
* 6 reads, 4 writes to a device with a queue depth of 2
* 4 writes to a device with a queue depth greater than 4
* 1 read
* 1 write
Results show that packetized SCSI improves throughput for all transfer sizes and workloads. The same results appear with the bus running at 320MB/sec.
THE FUTURE: PACKETIZED SCSI PLUS FIBRE CHANNEL
While SCSI controls most of the server and subsystem market, Fibre Channel is the new kid on the block. SCSI has a proven track record. Its been around a long time and is a very stable, compatible interface. However, Fibre Channel allows the direct attachment of many more devices over a greater distance. In an environment where this is the requirement, the preferred solution is to interconnect boxes of drives using Fibre Channel, keeping the interface within the box parallel SCSI. This solution preserves the investment in traditional SCSI while realizing the benefits of Fibre Channel.
However, parallel SCSI may still be inefficient at Fibre Channel data rates. In addition, translation is needed to take a Fibre Channel serial data stream and convert it into a parallel SCSI data stream. This is a complex operation requiring sophisticated algorithms. Packetized SCSI makes parallel SCSI more efficient for moving data on a parallel SCSI bus. It runs at all parallel data transfer rates, is completely backward compatible, and uses data structures that closely match those in Fibre Channel.
Adding packetized SCSI allows further benefit to the use of Fibre Channel. Running packetized SCSI on the parallel SCSI devices requires almost no translation of the Fibre Channel frame contents because the format and content of the data within the SCSI packets is nearly identical to that in the Fibre Channel frames. To preserve much of the existing physical and mechanical design, as well as provide an immediate step between current SCSI designs and full Fibre Channel designs:
* Move to Fibre Channel storage subsystems or network attach subsystems by putting a Fibre Channel on the subsystem
* Use existing parallel SCSI devices to minimize cost and maximize current investment by leaving the interconnect structure in place
* Create a seamless flow of data between Fibre Channel and parallel SCSI interconnects by adding the packetized SCSI protocol
Packetized SCSI can provide significant benefit over the traditional parallel SCSI. Performance improvements with packetized SCSI can be significant. In addition, packetized SCSI can be used in combination with the relatively new Fibre Channel interface to preserve the existing investment, as well as provide an immediate step in the transition from existing parallel SCSI to full Fibre Channel design.
George Penokie is the senior engineer of IBM Corporation (Rochester, MN).
|Printer friendly Cite/link Email Feedback|
|Title Annotation:||Technology Information|
|Publication:||Computer Technology Review|
|Date:||Jan 1, 2000|
|Previous Article:||UltraSCSI For Storage Lives On Despite The new Fibre Channel Kid On The Block.|
|Next Article:||Al Shuqart Remembers.|