Chapter 0

• Process
  • users submit jobs (programs and data) to operators
  • operators sort jobs manually (with similar needs) into batches
  • OS simply runs each job sequentially (no decision-making)
• Features
  • One job at a time
  • no user interaction
  • CPU is often idle, waiting for slow IO devices

• Overlaps CPU and I/O → ↑ CPU utilization
• Spooling: Simultaneous Peripheral Operations On-Line
  • It's only a concept ← implemented by interrupt
  • I/O is done with no CPU intervention
  • CPU just be notified when I/O is done

• interactive: provide direct communication between user and system
• users: share the CPU simultaneously

Feature	Batch	Multiprogramming	Time-Sharing
System Model	single user, single job	multiple prog.	multiple users, multiple prog.
Purpose	simple	resource utilization	interactive response time
OS feature	N/A	CPU scheduling, memory management	file system, virtual memory, sync, deadlock

• convenience and responsiveness -- GUI
• I/O devices
• lack of OS protection from users

• multiple CPUs communicate through shared memory
• purpose: ↑ throughput, economical, ↑ reliability
• SMP (Symmetric Multiprocessing System) -- most popular
  • each processor runs the same OS
  • require extensive sync to protect data integrity
• Asymmetric Multiprocessing System
  • each processor is assigned a specific task
  • one master CPU & multiple slave CPUs
  • for extremely large systems
• Multi-core Processors
  • on-chip communication is faster than between-chip communication
  • problem: different core have different cache
• Many-core Processors
  • hundreds or thousands of cores
  • e.g. GPU (Graphics Processing Unit)
• Memory Access Architecture
  • UMA (Uniform Memory Access): all processors share the same physical memory uniformly
    • identical access time
  • NUMA (Non-Uniform Memory Access): each processor has its own local memory, but can also access memory owned by other processors (higher latency)
    • non-uniform access time

• communicate through I/O bus or network
• easy to scale to large # of nodes
• purpose: resource sharing, load sharing, reliability
• architecture:
  • client-server model: easy to manage but bottleneck and single point of failure
  • peer-to-peer model: each node has equivalent capabilities and responsibilities
• Clustered Systems
  • multiple systems connected through a high-speed network (e.g. Infiniband)
• Supercomputers
  • performance measured in FLOPS (Floating Point Operations Per Second), not MIPS (Million Instructions Per Second)
• Heterogeneous Computing
  • different types of computing units (DSP, FPGA, GPU, CPU ...)
• HSA (Heterogeneous Systems Architecture)
  • unified memory address space for all computing units
  • eliminate data copying between different memories

Single core → Multi-core → Many-core (tightly) | (loosely) Cluster → Distributed