Straggler Step Time Impact

Problem

Given per-rank step times in milliseconds, a synchronous data-parallel step takes the slowest rank's time. Return [slowest_ms, slowdown], where slowdown = slowest / mean - 1 measures how much the straggler inflates the average.

The signature is straggler_step_time(rank_step_times_ms).

Hint

A single slow rank forces every other rank to wait at the synchronization barrier.

Test cases

case 1

in: [[100,100,100,120]]

expected: [120,0.1428571429]

case 2

in: [[50,50]]

expected: [50,0]

case 3

in: [[80,90,100]]

expected: [100,0.1111111111]

case 4

in: [[200]]

expected: [200,0]

case 5

in: [[10,20,30,40]]

expected: [40,0.6]

Starter codePython

def straggler_step_time(rank_step_times_ms):
    # Your code here
    pass

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