Please enable JavaScript.
Coggle requires JavaScript to display documents.
Parallel Processors (GPU (high data-parallel (momory latency -> thread…
Parallel Processors
GPU
video cards
3D graphics processing
high data-parallel
momory latency -> thread switching으로...
MIMD에 조금. 퍼포먼스는 낮음.
Difficulties
Partitioning
Coordination
Communications overhead
Scaling
Strong scaling :: problem size fixed
Weak scaling: problem size proportional to
number of processors
Instruction streams
SPMD : Single program multiple data
All processors execute the same
instruction at the same time
reduced control.
Simplifies synchronization
multithreading
Fine-grain multithreading
each cycle 마다 switch
Coarse-grain multithreading
long stall 때만 switch
Simultaneous Multithreading
Schedule instructions
independent threads
unit available -> execute
Tolerating cache-miss latency
core의 resourves share 증가.
vector proccessors : highly piplined function units.
to and from vecctor
ex) data and registers
lv,sc,addv,addvs
vector vs scalar
avoiding loops.
vector : variable vector width
pipelined and
arrayed functional units:
multimedia extensions: fixed width
Amdahl’s Law
parallel에는 적용불가.
Sharedmemory
SMP :Shared memory multiprocessor
message passing. -> 하나하나 private
Loosely Coupled Clusters
네트워크에 각각이 독립적
각각의 일에 맞춤.
but, 관리비용.... VM선호
low interconnect bandwidth
Parallel Benchmarks
traditional : fixed code and datas
Connection multiple computers
