Outline

The order (and to a lesser extent the content) are subject to change.

• Processors
  • who makes them, performance trends, heat and power replacing speed as main concerns
  • physics and economics of processor design
• Instruction execution sequence
  • basic von Neuman architecture, CISC
  • RISC
  • basic pipelined processor
  • stalls for data dependency and solutions
  • out of order completion
  • full out of order execution
  • superscalar processor
  • VLIW
  • vector pipelining
  • stalls for control dependency, branch prediction
  • interrupts and exceptions
  • cost modelling issues
• Memory hierarchies
  • latency and throughput, typical values
  • properties of an ideal memory hierarchy
  • locality - approximating the ideal
  • where locality comes from
  • effects of object-oriented programs on locality
  • effects of context switch on locality
  • basic design of caches, block loads
  • placement strategy, direct, associative set-associative, stride
  • replacement strategy
  • cache parameters and their dependence on design decisions
  • prefetching
  • writing cached data
  • instruction versus data caches
  • asymmetry of load and store
  • cost modelling the storage hierarchy
  • compulsory, capacity, conflict misses
  • exposing the cache to compilers
  • processors in memory
  • percolation
• The SIMD idea
  • system design
  • processor design
  • dynamic networks
  • static networks: path, torus, trees, hypercube
  • network properties
• Programming SIMD architectures
  • vector style
  • BMF style - map, reduce, scan
  • string search example
  • success stories
  • SPMD
• Shared-memory MIMD architectures
  • basic system design
  • why not a dynamic network
  • buses and serialization
  • cache coherence
  • snoopy protocols
  • table driven protocols
  • virtual shared memory/NUMA
  • clumps
  • dynamic network enhancements: CREW, CRCW, fluent
• Programming smMIMD architectures
  • PRAM
  • locks, semaphores
  • Cilk - C with fork, work stealing, race conditions
  • OpenMP
• Distributed-memory MIMD architectures
  • Origins -- IMPs, packet switching, store and forward, error model
  • TCP
  • static network topologies
  • virtual cutthrough, wormhole routing, deadlock avoidance, QoS
  • deterministic vs adaptive routing
  • switches, hierarchies
  • getting to the network: TCP/IP vs UDP vs kernel and user space NIC models
  • basic cluster design
  • economics of clusters
  • cluster networks: Enet, channel bonding, Gbit Enet, Myrinet, VIA
  • constellations
• Programming dmMIMD architectures
  • placement, data with computation
  • master-slave
  • message passing: MPI
  • BSP
• Multithreading
  • P-RISC
  • Design issues: critical path, fork/join, memory latency
• Dataflow
  • Static vs dynamic dataflow
  • eager vs lazy evaluation
  • using dataflow at coarse grain
• Other programming models
  • higher-order functional programming, string and graph reduction
  • Linda
  • actors and other object-based models
  • skeletons
• Hierarchy of models by architecture independence
• Hierarchy of systems by computational power
• Computational grids
  • motivation
  • design issues
  • variations: datacentric grids
• Quantum computers