Scaling Efficiency From Step Time

Problem

Compute data-parallel scaling efficiency from measured step times: efficiency = single_gpu_step_ms / (num_gpus * multi_gpu_step_ms). A value of 1.0 is perfect scaling, below 1.0 means communication or straggler overhead, and above 1.0 means superlinear speedup.

The signature is scaling_efficiency(single_gpu_step_ms, multi_gpu_step_ms, num_gpus).

Hint

Ideal multi-GPU time is single-GPU time divided by num_gpus.

Test cases

case 1

in: [100,15,8]

expected: 0.8333333333

case 2

in: [100,12.5,8]

expected: 1

case 3

in: [50,10,4]

expected: 1.25

case 4

in: [100,100,1]

expected: 1

Starter codePython

def scaling_efficiency(single_gpu_step_ms, multi_gpu_step_ms, num_gpus):
    # Your code here
    pass

4 test cases · run your code against them in the editor