Computer Performance Issues and Microprocessors

Lecture 2: Performance Issues

• The cost of computer systems continues to decrease dramatically, while performance and capacity increase equally.
• Today's laptops have computing power similar to an IBM mainframe 10-15 years ago.
• Processors (microprocessors) often have inexpensive cost, potentially disposable.
• Desktop applications requiring significant microprocessor-based system power include: image processing, 3D rendering, speech recognition, video conferencing, multimedia authoring, voice and video annotation of files, and simulation modeling.
• Businesses rely on powerful servers for transaction and database processing and support massive client/server networks that replace large mainframe systems.
• Cloud service providers utilize massive high-performance server banks to support high-volume, high-transaction-rate applications.

Microprocessor Speed

• Contemporary processors include techniques like: pipelining, branch prediction, superscalar execution, and data flow analysis, and speculative execution.
• Pipelining moves data/instructions through stages simultaneously, improving workflow.
• Branch prediction anticipates instruction sequences, streamlining processing.
• Superscalar execution allows multiple instructions per clock cycle using parallel pipelines.
• Data flow analysis prioritizes instructions based on dependencies for optimized scheduling.
• Speculative execution temporarily anticipates instructions ahead of execution, maximizing engine use.

Performance Balance

• Adjusting organization and architecture to compensate for varying component capabilities is key.
• Architectural examples include: increasing the number of bits retrieved simultaneously by widening data paths, reducing memory access frequency by adding complex caches between processor and main memory, and improving DRAM interface efficiency through schemes like on-chip caching.
• Increasing interconnect bandwidth (communication between processors and memory) helps with faster buses and a tiered bus structure.

I/O Device Data Rates

• A chart showing data rates (bits per second) for various input/output (I/O) devices.
• Examples include Ethernet modem, graphics displays, Wi-Fi modems, hard drives, optical drives, laser printers, scanners, mice, and keyboards.

Improvements in Chip Organization and Architecture

• Increasing processor speed, fundamentally, arises from shrinking logic gate sizes and packing more tightly (more gates), increasing clock rate.
• Signal propagation times reduce as circuit density increases.
• Increasing and speeding up on-chip caches improve performance.
• Revising processor organization and architecture leads to improved instruction execution speed and parallelization.

Problems with Clock Speed and Logic Density

• Power density increases directly with logic and clock speed (more transistors use more power).
• Heat dissipation is crucial.
• Speed (electron flow) is limited by factors like resistance and wire capacitance, increasing delays as components/wires shrink.
• As components shrink, interconnect resistance increases while capacitance also increases, making speed enhancement harder.
• Memory latency (memory speed) typically lags behind processor speed increases and limits overall performance.

Processor Trends

• A graph demonstrating the increasing trend of transistors (thousands), frequency (MHz), power (watts), and cores over time starting roughly around 1970 – showing exponential growth in transistors despite the limitations of processing power (power consumed).

Multicore

• Utilizing multiple processors on a single chip improves performance without increasing clock speed.
• Simplified processors are often advantageous over more complex processors.
• Larger caches are often used in dual processor setups to optimize processing efficiency.
• Multicore processors are more efficient in utilizing available cache resources, which are crucial for performance optimization.

Many Integrated Core (MIC) Graphics Processing Unit (GPU)

• A leap in performance resulting from advanced software in dealing with numerous cores.
• Homogeneous processors on a single chip are advantageous.
• GPU optimized cores are built for high-parallel operations, especially for graphics data.
• Typical use in plug-in graphics cards, encoding and rendering two- and three-dimensional graphics, as well as video processing.

Amdahl's Law

• A fundamental calculation related to potential speedup when employing multiple processors in programs.
• It highlights challenges in multi-core machine development.
• Software needs adaptations for highly parallel environments to maximize parallel processing power.
• It provides a generalized technique to evaluate and design computer system improvements.

Little's Law

• A straightforward formula applicable to systems in a stable state that exhibit no loss/leakage of elements.
• Applicable to queuing systems, demonstrating a relationship between item arrival rates and average wait times.
• It has limited assumptions to make it generally applicable.

Calculating the Mean

• Benchmarks compare systems by determining the average value of execution times using various (arithmetic, geometric, harmonic) mean calculations.
• These common means are needed to assess benchmarks when comparing systems' performance.

Benchmark Principles

• A benchmark program should meet important characteristics.
• It should be written in high-level languages for platform portability.
• It should be representative of typical programming uses to provide fair comparisons and demonstrate typical performance, such as in systems programming, numerical computing, or commercial programming.
• It should be amenable to easy measurement.
• It should have a broad distribution across platforms for greater applicability.

System Performance Evaluation Corporation (SPEC)

• SPEC is a benchmark suite consisting of various programs, commonly used to test computer performance in specific application areas.
• It represents a collection of industry-standard suites aimed at assessing system performance for benchmarking purposes and evaluating computer systems for comparative analysis and research purposes.

SPEC CPU2006

• A well-known benchmarking suite used as a standard test for measuring processor performance in applications that spend most of their time computing (rather than input/output).
• Consists of (17+12 programs) floating-point and integer programs written in C, C++, and Fortran.

Terms Used in SPEC Documentation

• Benchmark, system under test, reference machine, and base metric are terms critical to the methodology and interpretation of SPEC benchmarks.

SPEC Evaluation Flowchart

• This flowchart outlines the process used to evaluate computer system performance using the SPEC methodology.

SPEC CINT2006 Results

• A table presenting SPEC CINT2006 benchmark results (execution times) on two different computing platforms.

SPEC CPU2006 Floating-Point Benchmarks

• Further benchmark results for various floating-point applications.

This quiz focuses on performance issues related to modern computer systems and microprocessors. It covers advancements in computing power, applications relying on high-performance systems, and techniques that improve processor efficiency, such as pipelining and branch prediction. Test your knowledge on how these concepts apply to contemporary computing environments.