Ch 1.6-1.7 (HP1)

1.6 Performance

• Defining Performance
• Two different measures of performance
• Execution/Response Time: Time between the start and completion of a task
• 
• Throughput/Bandwidth: Total amount of work done in a given time
• Measuring Performance
• Program execution time is measured in seconds per program
• CPU Execution Time: Time the CPU spends computing for a specific task and does not include time spent waiting for I/O or running other programs as a CPU may work on more than one task at once
• User CPU Time: Time spent in the program
• System CPU Time: Amount of time running system calls
• Book uses system performance as CPU Time and performance to elapsed time on an unloaded system
• Clock Cycle/Period: Discrete time intervals a CPU works with
• 
• CPU Performance and Its Factors
• 
• Improve performance via decreasing # clock cycles
• Improve performance via decreasing clock cycle time
• Instruction Performance
• Another metric measures the performance in the amount of clock cycles required to run the program
• 
• Average Clock Cycles per Instruction: CPI
• The Classic CPU Performance Equation
• 

• Shows 3 different metrics for CPU time in terms of instruction count
• CPI can be obtained via software or hardware counters
• Instruction count depends on the architecture and not the implementation
• CPI depends on many things including memory design and processor structure and types of instructions that dominate execution
• Some CPUs vary the clock rate to save power
• Some instructions take < 1 CPI because things can be parallelized
• • 

• Recently we have reached the practical power limit for cooling commodity microprocessors
• Therefore the clock rate has been increasing slowly and companies are focusing more on the decrease of power consumption
• Energy usage is better measured joules than watts (joules/second)
• In CMOS transistors, primary energy consumption is dynamic energy
• Dynamic Energy: Energy used from switching between 0-1 
• Dynamic energy depends on the capacitive loading of each transistor and voltage applied
• Energy is proportional to Capacitive Load * Voltage^2
• This equation is the energy of a pulse: 0->1->0 or 1->0->1
• The energy of a single transition is
• 1/2 Capacitive Load * Voltage^2
• Power Single Transition=CV^2/2 * Frequency Switched
• Frequency switched is a function of clock rate
• Capacitive load per transistor is a function of both # transistors connected to output (fanout) and the technology used
• The reason why processor speeds used to grow much faster than power consumption is because power is a function of voltage^2 which was reduced at about 15% per generation