Bamboo Trimming Revisited: Simple Algorithms Can Do Well Too

Abstract

The bamboo trimming problem considers n bamboo with growth rates h 1 , h 2 , . . . , h n satisfying i h i = 1. During a given unit of time, each bamboo grows by h i , and then the bamboo-trimming algorithm gets to trim one of the bamboo back down to height zero. The goal is to minimize the height of the tallest bamboo, also known as the backlog. The bamboo trimming problem is closely related to many scheduling problems, and can be viewed as a variation of the widely-studied fixed-rate cup game, but with constant-factor resource augmentation.Past work has given sophisticated pinwheel algorithms that achieve the optimal backlog of 2 in the bamboo trimming problem. It has remained an open question, however, whether there exists a simple algorithm with the same guarantee-recent work has devoted considerable theoretical and experimental effort to answering this question. Two algorithms, in particular, have appeared as natural candidates: the Reduce-Max algorithm (which always cuts the tallest bamboo), and the Reduce-Fastest(x) algorithm (which cuts the fastest-growing bamboo out of those that have at least some height x). It is conjectured that Reduce-Max and Reduce-Fastest(1) both achieve backlog 2.This paper improves the bounds for both Reduce-Fastest and Reduce-Max. Among other results, we show that the exact optimal backlog for Reduce-Fastest(x) is x + 1 for all x ≥ 2 (this proves a conjecture of D'Emidio, Di Stefano, and Navarra in the case of x = 2), and we show that Reduce-Fastest(1) does not achieve backlog 2 (this disproves a conjecture of D'Emidio, Di Stefano, and Navarra).Finally, we show that there is a different algorithm, which we call the Deadline-Driven Strategy, that is both very simple and achieves the optimal backlog of 2. This resolves the question as to whether there exists a simple worst-case optimal algorithm for the bamboo-trimming problem.

Petals: Collaborative Inference and Fine-tuning of Large Models

“…Recent works demonstrate that model parallelism can be used for cost-efficient pre-training of LLMs by pooling together idle GPU devices (Athlur et al, 2022;Wang et al, 2022;Kuszmaul, 2022;Yuan et al, 2022;Ryabinin et al, 2023).…”