1. Gary MarsdenSlide 1University of Cape Town Computer Architecture – Introduction Andrew Hutchinson & Gary Marsden (me) ( gaz@cs.uct.ac.za ) 2005

2. Gary MarsdenSlide 2University of Cape Town The Grand Scheme  Abstractions  Role of Performance  Language of the Machine  Arithmetic  [3] Performance Issues  (6) Processor: Datapath & Control  (3) Pipelining  (4) Memory hierarchy  (4) Peripherals CS1 CS2

3. Gary MarsdenSlide 3University of Cape Town Review  Chapter 1: Abstractions –HLL -> Assembly -> Binary –Application (system s/w (hardware)) –5 classic components Input, output, memory, datapath, control –Processor gets inst. & data from memory –Input for placing items in memory –Output reads from memory –Control co-ordinates –Abstraction / Interfaces

4. Gary MarsdenSlide 4University of Cape Town Review II  Chapter 2: The Role of Performance  CPU time =  Popular measures –MIPS Depends on inst. Set | Varies between progs. | Can vary inversely with performance –GFLOPS Better, perhaps, as instructions are similar FLOPS can differ across machines Instruction Count x CPI Clock Rate

5. Gary MarsdenSlide 5University of Cape Town Review III  Chapter 3 - Machine Language –R type and I type (distinguished by first field) –Registers –Address modes Register, base, immediate, PC relative –P1: Simplicity favours regularity –P2: Smaller is faster –P3: Good design demands compromise –P4: Make the common case fast  Chapter 4 - Arithmetic –ALU construction

6. Gary MarsdenSlide 6University of Cape Town New Chapter 4 - Performance  What is a fast computer? –Clock speed? –Responsiveness? –Data processing?  It depends…

7. Gary MarsdenSlide 7University of Cape Town What is performance?  There are so many elements to a computer system that it is pretty much impossible to determine, in advance, the performance level of a system  Computer vendors have their own idea of ‘performance’ –Macintosh users have the ‘bounce’ test  Performance means different things to different people –747 or F-15

8. Gary MarsdenSlide 8University of Cape Town Two key measures  Users of personal computers are interested in ‘response time’ –Also called execution time –F-15  Data centre managers are interested in ‘throughput time’ –Total work done in a given time –747  This implies at least two different performance ‘metrics’

9. Gary MarsdenSlide 9University of Cape Town Response time  As we are looking at microprocessor design, response time is of primary interest  Start by considering execution time –Longer is bad; shorter is good  So we can say that

10. Gary MarsdenSlide 10University of Cape Town What is time?  “Time is an illusion; lunch time doubly so”  We kind of cheated when we said “time”  Time can be thought of in two ways –Wall time / Response time /Elapsed time: all mean the number of seconds passed since the task was started until it ends –CPU time: In multi-user systems, it makes more sense to count the time the CPU spends on that user’s task (‘system’ time confuses issue) Try the Unix ‘time’ command

11. Gary MarsdenSlide 11University of Cape Town CPU performance  In personal computers, most people focus on clock speed - e.g. 4 Ghz processor –4 GHz is the clock rate; 0.25 ns is the cycle length  CPU execution time = Clock cycles for a program  So increasing clock speed, or decreasing the number of cycles for a program will improve performance  We need to think a bit more about ‘cycles for a program’ Clock Rate

12. Gary MarsdenSlide 12University of Cape Town Program length  The number of cycles require for a program will depend on the number of instructions contained in a program and the number of cycles it takes to execute each of these instructions –We average this to CPI (Cycles per Instruction)  So… –CPU clock cycles = Number of instructions x CPI  Therefore –CPU execution time = Number of instructions x CPI Clock Rate

13. Gary MarsdenSlide 13University of Cape Town A note on CPI  Different instructions take different numbers of clock cycles to complete –A lot more on this later  By using hardware and software monitoring tools, one can calculate a sufficiently useful CPI value

14. Gary MarsdenSlide 14University of Cape Town So how can I make my program faster?  Algorithm: Instruction count (and possibly CPI) –Algorithm choice affects number and type of instructions  Language: Instruction count and CPI –Some languages require more statements per expression and may require more high-cycle instructions  Compiler: Instruction count and CPI –Compilers can have a huge effect on both these measures; too complex to deal with here  Processor Instruction Set: CPI, clock rate & instruction count –We will get in to this in the next chapter

15. Gary MarsdenSlide 15University of Cape Town Comparative Performance  So we know the elements, but how do we compare systems?  For situations where there is only one application to run, this is straightforward  Most of us, however, run multiple applications  Could test the application on different platforms, but only if it is available for target platforms  Alternatively create a simple application to create an idealised usage –The Benchmark!

16. Gary MarsdenSlide 16University of Cape Town Benchmarks  There are many of these created to indicate performance for different types of task  A good place to look is –www.spec.org  However, some compilers are created to optimise for specific benchmarks  Hence the need for reproducability

17. Gary MarsdenSlide 17University of Cape Town You may have head of MIPS?  MIPS (Millions of Instructions Per Second) was, for a long time, the standard benchmark, but has now been replaced  On the plus side, bigger numbers usually mean faster computers  On the down side –Different instruction sets do different things –MIPS ratings vary per program (different MIPS for same computer) –MIPS can vary inversely with performance.

18. Gary MarsdenSlide 18University of Cape Town Some performance laws  Moore’s law: “The complexity of an integrated circuit will double every 18 months” –What are the implications of that?  Amdahl’s Law: Even dramatic improvements in part of a task will not have significant effects on the overall task (diminishing returns) –Implication is that the common task should be fast