Download Computer System Design - Computer Architecture | ECEN 4243 and more Assignments Computer Architecture and Organization in PDF only on Docsity!
COMPUTER SYSTEM DESIGN
Speed & Cost. Speed measure: how fast you get the service you desire (includes hard- ware and software)
Cost measure: acquisition, training and maintenance costs (includes hardware and soft- ware)
Speed/Cost Trade-off. Migrating functionality to hardware:
Increases speed but increases cost
Migrating functionality to software:
decreases cost but decreases speed
moral: unless functionality significantly improves performance (i.e. speed), don’t put it in hardware (to reduce cost).
Reduced Instruction Set Computer (RISC) vs. Complex Instruction Set Computer (CISC)
- No sharp dividing line! - Not obvious which is best for reducing program execution time! - RISC approach does seem to reduce hardware cost (at least of processor)
RISC processors dominate modern processor design since they give good performance at reduced hardware cost.
RISC CISC
- smaller number of instructions - larger number of instructions - simple fast instructions • complicated slower instructions - more instructions in program - less instructions in program
Performance Driven Design
Moral: computer design is application dependent
Application for study in this class: General purpose computing for programs written in a high level language (HLL).
Performance Measurement General definition:
which we want to maximize. Alternatively, we can choose to minimize the inverse of speed:
CPU time can be broken into three major components (see p. 249):
Service
desired
by user
performance
requirements
instruction
set
complexity
computer
design
instruction
set
operating system,
compilers, linkers,
user programs
Software
Central Processing
Unit (CPU), memory,
input/output (I/O)
Hardware
We will study what typi- cal software requires.
We will cover a detailed example of a single instruction set and hardware design.
speed
programs executed
time
CPUtime
time
programs executed
CPUtime
instructions
programs executed
clock cycles
instruction
time
clock cycles
= ^
= ( IC) ( CPI) ( T)
- CPI depends on what program is executed. - CPI is improved most by improving CPI i for the instructions with the largest IC i.
Benchmarks. Performance depends on what program is being executed. In order to com- pare the performance of different computer designs, we really should compare designs running the same “benchmark” programs.
But the different computers will have different instruction sets, so what do we mean by the same program? We mean programs written in high level languages like FORTRAN, C, or COBOL.
Benchmarks measure performance of the compiler + hardware combination.
Commonly used set of benchmarks: fig. 4.5, p.
Results are usually reported individually for each of the benchmark programs. This is more meaningful than using average performance.
For example, the designers of CPU2, above, get a low average CPU time by tuning their design to run very fast on prog3, but CPU1 might be the better general purpose machine.
Amdahl’s Law. Suppose a part of the computer is changed to be faster than originally designed.
Suppose further that a FracTimeenhanced fraction of the time on the old design, the new
design can use the new hardware that has a speedup of Speedupenhanced. The rest of the
time the new design uses the same hardware with the same speed. Then,
prog
CPU time
prog2 prog3 prog4 prog
CPU
CPU
Speedup overall
Speed new
Speed old
CPUtime old
CPUtime new
CPUtime old FracTime enhanced ×CPUtime old
+( 1 - FracTime enhanced) ×CPUtime old
which is Amdahl’s law (p. 267). Amdahl’s law shows that speeding up the hardware is only effective when the improved hardware can be used a large fraction of the time. This is equivalent to improving the instructions with the highest instruction count or frequency.
CPUtime new FracTime enhanced
CPUtime old
Speedup enhanced
×--------------------------------------------
+( 1 - FracTime enhanced) ×CPUtime old
Speedup overall
FracTime enhanced
Speedup enhanced
-------------------------------------------- 1 - FracTime
+ enhanced
=^ ------------------------------------------------------------------------------------------------------
