IBM 990 manual Processor Unit design, Dual External Time Reference, Superscalar processor

Draft Document for Review April 7, 2004 6:15 pm 6947ch02.fm

Data transfer between the CEC memory and attached I/O devices or CPCs is done through the Memory Bus Adapter. The physical path includes the Channel card (except for STI connected CPCs), the Self-Timed Interconnect bus and possibly a STI extender card, the Storage Control, and the Storage Data chips.

More detailed information about I/O connectivity and channel types can be found in “I/O subsystem” on page 71.

Dual External Time Reference

The optional ETR connections, although not part of the book design, are found adjacent to the books on the opposite side of the CEC board. The z990 servers implement an Enhanced ETR Attachment Facility (EEAF) designed to provide a dual External Time Reference (ETR) attachment facility. Two ETR cards are automatically shipped when Coupling Links are ordered and provide a dual path interface to the IBM Sysplex Timers, which are used for timing synchronization between systems in a Sysplex environment. This allows continued operation even if a single ETR card fails. This redundant design also allows concurrent maintenance.

2.2.3 Processor Unit design

Each PU is optimized to meet the demands of new e-business workloads, without compromising the performance characteristics of traditional workloads. The PUs in the z990 have a superscalar design.

Superscalar processor

A scalar processor is a processor that is based on a single issue architecture, which means that only a single instruction is executed at a time. A superscalar processor allows concurrent execution of instruction by adding additional resources onto the microprocessor to achieve more parallelism by creating multiple pipelines, each working on their own set of instructions.

A superscalar processor is based on a multi-issue architecture. In such a processor, where multiple instructions can be executed at each cycle, a higher level of complexity is reached because an operation in one pipeline may depend on data in another pipeline. A superscalar design therefore demands careful consideration of which instruction sequences can successfully operate in a multi-pipeline environment.

As an example, consider the following: if the branch prediction logic of the microprocessor makes the wrong prediction, it might be necessary to remove all instructions in the parallel pipelines also (refer to “Processor Branch History Table (BHT)” on page 44 for more details).

There are challenges in creating an efficient superscalar processor. The superscalar design of the z990 PU has made big strides in avoiding address generation interlock situations. Instructions requiring to get information from memory locations may suffer multi cycle delays to get the memory content. The superscalar design of the z990 PU tries to overcome these delays by continuing to execute (single cycle) instructions that do not cause delays. The technique used is called “out-of-order operand fetching”. This means that some instructions in the instruction stream are already underway, while earlier instructions in the instruction stream that cause delays due to storage references, take longer. Eventually the delayed instructions catch up with the already fetched instructions and all are executed in the designated order. The z990 PU gets much of its superscalar performance benefits from avoiding address generation interlocks.

It is not only the processor that contributes to the capability of successful execution of instructions in parallel. Given a superscalar design, compilers and interpreters must create code that benefit optimally from the particular superscalar processor implementation. Work is

Chapter 2. System structure and design 41