應用於平行遊戲樹搜尋之高效能計算環境之碎片感知排程

Abstract

在高效能計算系統中，會根據一些排程方法來分配資源給使用者，但這些排程方法在分配的過程中，通常都會產生一些資源碎片，本文將探討使用資源碎片的策略和技術。我們提出以實現在平行蒙地卡羅樹狀搜尋（MCTS）的輕量級單核心任務，去適應不穩定的碎片資源環境。本篇論文使用我們實驗室開發的一套系統來收集算圖農場上的資源碎片，並利用這些資源碎片來執行MCTS的工作。這套系統傾向於將重要的樹節點分配到可靠的計算節點上運行，來確保此樹節點優先被探索和順利完成。我們的實驗結果成功的將MCTS計算融合在碎片資源環境中，並且可在工作產量和工作成功率之間做權衡。

Due to the complex scheduling criteria, resource fragmentation often occurs in high performance computing systems. This paper addresses both of the policy and technical issues of using unstable resources formed by fragmentation. We propose that the parallel Monte-Carlo tree search (MCTS) can be implemented as lightweight single-core tasks to adapt to the unstable fractured resources. We develop a system to collect idle resources in a high performance render farm and make use of them to execute MCTS tasks. The system tends to assign promising tree nodes to reliable and responsive processors to confirm best-first search. Our work demonstrates a successful integration in which the MCTS computation gains significant resources without interfering with the render farm tasks.