CENG 446

1
CENG 446 Advanced Computer Architectures

• Dr. Brian T. Hemmelman
• Chapter 1 Slides

2
The Many Faces of Computers

• Desktops Generally for a single user running a
  wide variety of software.
• Servers Handle many tasks for many users
  (scientific, database, Web).
• Supercomputers Typically designed for intensive
  scientific and engineering modeling and
  computation.
• Datacenters Massive collections of processors
  or servers to handle search engines or
  e-commerce.
• Embedded Computers Contained within or are a
  part of other systems that perform a particular
  function (traffic light, automobile, microwave,
  smartphone, fly-by-light avionics system, etc.)

3
The Pervasiveness of Computers
4
Whats Going On Behind Your Program

• There are multiple layers of activity taking
  place for a computer to perform a particular
  task.
• Application Software A particular program
  designed to accomplish a specific task.
• Systems Software Software that provides
  services that are commonly useful.
• Operating System Supervising program that
  manages the resources of a computer.
• Compilers Translate a program written in a
  high-level language such as C or Java into
  instructions that the hardware can execute.
• Hardware The physical electronics and circuitry
  that perform the actual calculations or
  operations.

5
Hardware/Software Hierarchy
6
Bridging the Hardware/Software Interface

• The actual hardware uses voltage signals to
  represent information
• Boolean Logic Variable True/False
• Data Collections of 1s and 0s (bits)
• Computers must be told how to manipulate the
  information through instructions.
• Instructions are collections of 1s and 0s that
  force the hardware to do something with other
  collections of 1s and 0s.

7
Bridging the Hardware/Software Interface

• The binary instructions that are actually
  executed by the circuitry are called machine
  language or op codes.
• Creating all the binary instructions directly is
  cumbersome and tedious so helper programs were
  created to translate a symbolic notation of the
  desired operation into the necessary bit pattern
  for the machine.
• These programs are called assemblers and the
  symbolic notation is called assembly language.
• Eventually, assemblers were used to write more
  powerful translators that allowed a user to focus
  on the task algorithm instead of the circuitry
  dependent operations.
• These programs are the compilers, and the
  algorithmic codes are written in a high-level
  programming language.

8
The Journey of a Programmers Wish Into Physical
Reality
9
Advantages of High-Level Programming Languages

• The programmer focuses on the algorithm and
  describes it in a language more natural to
  humans.
• The programmer increases productivity as fewer
  lines of code are needed to describe how to
  execute the task. The process of converting the
  algorithm into machine language is automated by
  experts in translation. The programmer can focus
  on becoming an algorithm expert instead of an
  everything expert.
• Programs and algorithms can be designed that are
  largely independent of the specific processor or
  circuitry on which they will execute.
• These three advantages are so strong that today
  little programming is done in assembly language.

10
The Five Classic Computer Components

• Input
• Output
• Memory
• Datapath
• Control

11
Input/Output Examples

• Keyboard
• Mouse
• Analog-to-Digital Converter
• Monitor
• Network Connection
• Pulse-Width-Modulation (PWM) Signal

12
Memory

• External memory is almost always made of DRAM
  chips.

13
Memory

• Internal memory is usually made of SRAM memory
  cells.

14
More Memory

• Memory can also be categorized as volatile or
  nonvolatile.
• Volatile memory Only stores data when power is
  on (e.g. DRAM or cache).
• Nonvolative memory Data integrity is maintained
  even if no power (e.g. hard drive, FLASH, DVD).
• We could also distinguish main memory from
  secondary memory
• Main memory Volatile, where the program runs
  and its data is updated.
• Secondary memory Nonvolatile, where program and
  data are stored between runs.

15
Datapath and Control (The Processor or CPU)

• Datapath Performs the arithmetic operations.
• Control Tells the datapath, memory, and I/O
  devices what to do.

16
Changing Technology

• The increase in memory capacities and processor
  clock speeds have been incredible.

17
Changing Technology

• Computer performance has likewise been
  continuously increasing.

18
Power - The Limiting Factor
Power consumption is directly proportional to the
clock rate. Trying to increase computer
performance by only increasing the clock rate is
no longer feasible as heat dissipation becomes a
limiting factor. Hence the push towards
multicore processor chips running at somewhat
slower clock speeds.
19
Computer Performance

• So how do we know if one computer is better than
  a different computer?
• The answer, unfortunately, is not a simple one.
• Computer performance is largely application
  specific. Different applications have different
  needs and objectives, so no one performance
  measure is automatically The Best.
• Clock frequency used to be a way to get an easy
  measure of performance, but this is far to simple
  an approach for most systems today.

20
Defining Performance

• Books example Different airplanes with different
  ranges and speeds.
• Best performance could be defined in terms of
  greatest range, fastest cruising speed, moving
  the most passengers the quickest, etc.

21
Defining Performance

• For desktops, laptops, supercomputers, and
  embedded computer you are primarily interested in
  response time (execution time).
• Response time The total time required for the
  computer to complete a task, including disk
  accesses, memory accesses, I/O activities,
  operating system overhead, CPU execution time,
  and so on.

22
Comparing Performance

• Therefore, if comparing two computers X and y we
  could concludePerformanceX gt PerformanceY ifE
  xecution timeY gt Execution timeX
• We can then also compute ratios of performance as

23
CPU Performance

• CPU Execution Time The time CPU itself spends
  computing a particular task (does not include
  time spent waiting for I/O or running other
  programs).
• We could look at this from the simple perspective
  of the number of clock cycles it takes the CPU to
  complete the task. Thus,

CPU execution time for TaskA CPU clock cycles
for TaskA ? Clock cycle time CPU execution time
for TaskA (CPU clock cycles for TaskA)/(Clock
rate)

• Improving performance can then be achieved by
  reducing clock cycles required for the task
  (perhaps by more powerful instructions and hence
  more complicated circuitry) or decreasing the
  clock cycle time (increasing clock frequency).
• Decreasing clock cycle time tends to increase
  power consumption, and decreasing the clock
  cycles needed tends to increase clock cycle time.

24
Instruction Performance

• Addressing clock frequency, the power wall, and
  heat dissipation is a complete problem onto
  itself that will not specifically be covered in
  this class.
• We can however look into decreasing CPU clock
  cycles to complete a task.

CPU clock cycles (Instructions for TaskA) ?
(Average clock cycles/instruction)

• CPI (clock cycles per instruction) The average
  number of clock cycles each instruction takes to
  execute.
• Substituting this expression into the equation on
  the previous slide we obtain

25
The Classic CPU Performance Equation

• Our challenge as hardware designers is to
  optimize the balance between the instruction
  count and the CPI.
• How efficiently do we design the digital
  circuitry? Can we create different instructions
  that accomplish more without changing clock cycle
  time?
• Overall performance will also be affected by how
  well compilers utilize the instructions available
  in the instruction set implemented in hardware.

26
Example, Page 34
Suppose we have two implementations of the same
instruction set architecture. Computer A has a
clock cycle time of 250 ps and a CPI of 2.0 for
some program, and computer B has a clock cycle
time of 500 ps and a CPI of 1.2 for the same
program. Which computer is faster for this
program and by how much? Both computers need to
execute the same number of instructions, I, as
they are running the same program.
CPUA time I ? 2.0 ? 250 ps 500 ? I ps CPUB
time I ? 1.2 ? 500 ps 600 ? I ps
(CPUA performance)/(CPUB performance) (CPUB
time)/(CPUA time) (600I/500I) 1.2
CPUA will run 1.2 times faster.
27
Example, Page 35
A compiler designer is trying to decide between
two code sequences for a particular computer.
The hardware designers have supplied the
following facts
Instruction Class A Instruction Class B Instruction Class C
CPI 1 2 3
For a particular high-level language statement,
the compiler writer is considering two code
sequences that require the following instruction
counts
of Class A of Class B of Class C
Sequence 1 2 1 2
Sequence 2 4 1 1
Which code sequence executes the most
instructions? Which will be faster? What is the
CPI for each sequence?
28
Example, Page 35
The instruction count is trivial. Sequence 1
uses five instructions and Sequence 2 uses six
instructions. Which sequence is faster is
determined by calculating total CPU clock
cycles Sequence 1 clock cycles (2 ? 1) (1 ?
2) (2 ? 3) 10 clock cycles Sequence 2 clock
cycles (4 ? 1) (1 ? 2) (1 ? 3) 9 clock
cycles The CPI for each sequence is easily
computed as CPI (CPU clock cycles)/(Instruction
Count) CPI Sequence 1 (10 clock cycles)/(5
instructions) 2.0 CPI Sequence 2 (9 clock
cycles)/(6 instructions) 1.5
Conclusion Make the common case fast!
29
Case Study SPEC Benchmark for AMD Opteron X4
Model 2356
The integer portion of the benchmark, CINT2006,
is summarized in this table.
Those tasks with a CPI above 1.09 have a higher
CPI due to high cache miss rates, i.e. memory
access is slowing them down!