CMU Computer Systems: Thread-Level Parallelism

Exploiting parallel execution

• Use threads to deal with I/O delays
• Multi-core/Hyperthreaded CPUs offer another opportunity
  • Spread work over threads executing in parallel
  • Happens automatically, if many independent tasks
  • Can also write code to make one big task go faster

Typical Multicore Processor

Out-of-Order Processor Structure

Hyperthreading Implementation

Characterizing Parallel Program Performance

• p processor cores, T_k is the running time using k cores
• Speedup: $S_p= T_1 /T_p$
  • $S_p$ is relative speedup if $T_1$ is running rime of parallel version of the code running on 1 core
  • $S_p$ is absolute speedup if $T_1$ is running time of sequential version of code running on 1 core
  • Absolute speedup is a much truer measure of the benefits of parallelism
• Efficiency: $E_p=S_p/p = T_1/(pT_p)$
  • Reported as a percentage in the range (0, 100]
  • Measures the overhead due to parallelization

Amdahl's Law

• Captures the difficulty of using parallelism to speed things up
• Overall problem
  • T Total sequential time required
  • p Fraction of total that can be sped up ($0\leq p \leq 1$)
  • k Speedup factor
• Resulting Performance
  • $T_k=pT/k+(1−p)T$
    • Portion which can be sped up runs k times faster
    • Portion which cannot be sped up stays the same
  • Least possible running time
    • $k = ∞$
    • $T_∞=(1−p)T$

Experience with Parallel Partitioning

• Could not obtain speedup
• Speculate: Too much data copying
  • Could not do everything within source array
  • Set up temporary space for reassembling partition

Memory Consistency

• What are the possible values printed
  • Depends on memory consistency model
  • Abstract model of how hardware handles concurrent accesses
• Sequential consistency
  • Overall effect consistent with each individual thread
  • Otherwise, arbitrary interleaving

Snoopy Caches

• Tag each cache block with state
  • Invalid Cannot use value
  • Shared Readable copy
  • Exclusive Writeable copy