Lecture 3: Performance, The Power Wall, and Benchmarking (CSCI 4350)

Vocabulary flashcards covering key concepts from the lecture slides on performance, the power wall, and benchmarking.

Performance

The ability of a computer system to complete work promptly; in practice appears as execution time and throughput; time is the absolute metric.

Execution Time

The total wall-clock time for a program to complete its task, including all overheads (disk I/O, OS activity, etc.).

Throughput

The rate at which work is completed (e.g., tasks per unit time); often expressed as capacity × speed or as a completion rate.

Elapsed Time

Elapsed time equals the real-world time from start to finish of a program, including system activity; another name for execution time.

CPU Time

The portion of elapsed time during which the CPU is actively executing instructions for the program (user time plus system time).

User CPU Time

Time the CPU spends executing the program’s own instructions.

System CPU Time

Time the CPU spends executing operating-system or other privileged tasks on behalf of the program.

N(instr) / Instruction Count

The number of instructions executed by a program (not simply the number of source lines).

N(cycles) / Ncyc

The total number of CPU clock cycles required to execute a program.

CPI (Cycles Per Instruction)

Average number of clock cycles needed to execute each instruction; depends on algorithm, compiler, architecture, and implementation.

IPC (Instructions Per Cycle)

The inverse of CPI; IPC = 1/CPI. Higher IPC means more instructions completed per clock cycle.

f_clk / Clock Frequency

The number of clock cycles per second; measured in hertz (Hz). fclk = 1 / Tclk.

T_clk / Clock Period

The duration of one clock cycle; the reciprocal of clock frequency: Tclk = 1 / fclk.

Classic CPU Performance Equation

tcpu = NI × CPI × Tclk = NI × CPI / fclk; time to execute a program depends on instruction count, CPI, and clock rate.

MIPS

Millions of Instructions Per Second; a performance metric often computed as MIPS = fclk(MHz) / CPI (or NI / tcpu in millions of instructions per second).

SPEC CPU 2017

Standardized benchmark suite from SPEC to measure CPU performance (without I/O/graphics); includes multiple sub-suites for integer and floating-point, single- and multi-threaded workloads.

SPECspeed Integer 2017

SPEC speed suite focusing on integer single-thread performance.

SPECspeed Floating-Point 2017

SPEC speed suite focusing on floating-point single-thread performance.

SPECrate Integer 2017

SPEC rate suite focusing on integer multi-threaded throughput.

SPECrate Floating-Point 2017

SPEC rate suite focusing on floating-point multi-threaded throughput.

SPECratio

The ratio t(Ref) / t(SUT) for a SPEC benchmark; used to compute the overall SPEC score as a geometric mean of r-scores.

P_active / Power Dissipation

Active power in CMOS circuits, proportional to CL × V^2 × fclk (with switching activity). Higher frequency and voltage increase power.

Dennard Scaling

Scaling principle where transistor features scale down with voltage to keep power density manageable; real-world limits have reduced its effectiveness.

DFS / Dynamic Frequency Scaling

Technique to adjust clock frequency (and voltage) on demand to save power, activating only parts of the chip as needed.

Power Wall

Phenomenon where increases in clock rate yield rising power and heat, limiting sustained gains in single-thread performance; drives a shift to throughput via parallelism.

Multicore / Throughput

Increase system throughput by adding more cores; requires parallel programming and introduces scheduling and communication overhead (Amdahl’s Law).

Amdahl’s Law

Speedup of a program with multiple processors is limited by the serial portion of the program; overall speedup ≤ 1 / (serialfraction + (parallelfraction / numberofprocessors)).

Benchmark

A set of programs used to measure and compare system performance.

Benchmark Suite

A collection of standardized benchmarks (e.g., SPEC) used to characterize performance across workloads and domains.