Lecture Notes: Computer Architecture

Introduction

• Topics: Memory Ordering and Cache Coherence
• Importance: Relevant for multiprocessor and multicore systems
• Recap: Difficulty in parallel programming; trade-off between performance and correctness

Memory Ordering

Key Points

• Concept: Important for correct parallel programming in multiprocessor systems
• Trade-offs: Performance and correctness are at odds
• Sequential Consistency: Proposed by Lamport, ensures all processors see the same order of operations to memory
• Memory Consistency: Ordering of all memory operations across processors
• Coherence vs Consistency:
  • Coherence: Ordering of operations to the same memory location
  • Consistency: Global ordering of operations to all locations

Challenges

• Program Correctness: Maintaining correct execution in multiprocessors
• Debugging: Difficult without predictable order of operations

Models

• Sequential Consistency: Simplifies programmer's job, but costly in terms of hardware design
• Weak Consistency Models: Compromise made to reduce performance overhead while maintaining correctness
  • Release Consistency: Allows more parallel execution but is harder to implement

Cache Coherence

Key Points

• Problem: Ensuring all processors see the most recent value of cached memory
• Hardware Solutions: Preferable over software-only solutions due to complexity
• Invalidation Protocol: Simple approach where processors invalidate other copies when a cache block is written

Protocols

• MSI Protocol: Modified, Shared, Invalid states
  • Extend to MESI: Adds Exclusive state for optimization
  • MOESI Protocol: Optimizes further by adding Owner state

Implementation

• Snoopy Cache: Uses a shared bus for coherence, good for small systems
• Directory-Based Coherence: Scalable solution for large systems
  • Tradeoffs: More complex but avoids excessive broadcasting

Advanced Topics

• LazyPIM (Lazy Processing-In-Memory): Specific technique to handle coherence in near-memory computation environments
• Token Coherence: Combines benefits of Snoopy and Directory approaches

Exam Preparation

• Review sample exam questions
• Understand implementation challenges and trade-offs of different models

Conclusion

• Importance of finding balance between hardware complexity and programmer simplicity
• The lecture emphasized understanding coherence and consistency as fundamental for parallel computing in multicore systems.