Pipeline Parallellism with Controllable Memory
Pipeline Parallelism with Controllable Memory creates a framework on designing pipeline schedules and uses the framework to find memory optimal efficient schedules.
From our findings, we need approximately 1/3 memory under ideal conditions (F, B and W have same runtime), and 1/2 memory to create zero bubble schedule in realistic scenarios (with the necessary condition being W + 2B ≥ 2F and W + 2F ≥ 2B ).
Check out our paper at Arxiv.
| Method | 1F1B | V-Min | V-Half | V-ZB |
|---|---|---|---|---|
| Bubble Rate (assuming T_F=T_B=T_W) |
~ p/m | ~ 2p/3m | ~ p/ 2m | 0 |
| Activation Memory (by #micro-batch) |
p | (p+4)//3 | (p+2)//2 | p |
Bubble Rate here is calculated as 1 - (F+B+W)*m / longest_stage_time.