IOS: Inter-Operator Scheduler for CNN Acceleration

https://arxiv.org/abs/2011.01302

• Executive summary

  • Motivation: sequential execution --> under-utilization problem

• Inter-Operator Scheduler

  • Inter-operator parallelism

  • Dynamic programming --> optimal schedule

  • 1.1-1.5x speedup

• Efficient deployment of CNNs is important

  • Is CNN inference in current DL libraries well utilizing underlying hardware?

• Motivation for Inter-Operator Parallelization

  • More small convs in CNN design

  • GPU peak performance increased

  • Intra- and inter-operator parallelization

    • Sequential execution: Intra-operator Parallelization: Device under-utilization (small op & opwerful GPU)

    • Inter-Op Parallel Execution: better device utilization

• Background: wavefront schedule policy

  • Execute all available operators stage by stage

  • A better schedule

    • Put op to saturated stage: marginal benefit

    • Under-utilization problem

    • Wavefront schedule policy is sub-optimal

• Inter-operator scheduler (IOS)

  • General idea: explore the schedule space exhausitvely

  • Challenge: the number of schedules is exp in the number of operators

    • Prohibitive to enumerate

  • Observation 1: optimal schedule for a subgraph can be reused

    • Key idea: dynamic programming

  • Observation 2: the width of the computation graph is usually small (max number of parallelizable operators)

    • Key result: time complexity is only exponential in the width

  • 
  • Parallelization strategy selection

    • Concurrent execution --> multi-GPU kernel at the same time

    • Operator merge --> merged convolution, usually better performance

    • Profile & select

  • Last stage candidates

    • S' can be the last stage of S <--> there is no edge from S' to S - S'

  • Transition graph and time complexity

  • Methodology

    • Benchmarks

      • Inception V3, SqueezeNet, Randwire, NasNet

    • Baselines: state-of-the-art frameworks, different schedules on IOS Runtime

    • Environment: NVIDIA V100, Cuda, cuDNN

  • More active warps improve utilization