使用機器學習方法預測程序執行時間

Abstract

雲端運算可以在本地機器計算資源不足時提供計算資源。為了幫助 系統排程器決定一個程序該在本地端執行或者在雲端上執行,本論文 提供一個程序執行時間預測方法,使得排程器可根據預估的執行時間 決定要在本地或雲端執行此程序。本論文提出一個機器學習 (支持向量 回歸) 的方法,使用歷史資訊來建立預測模型。我們收集了程序執行的 歷史檔、系統負載資訊、檔案系統資訊,來預測重覆執行的高計算用 量程序執行時間。此方法實驗評估可達 20% 相對平均執行時間的平均 誤差,預測高計算用量開放原始碼程式的執行時間俱可行性。

Cloud computing is widely used in on-demand computing in recent years. A local computing cluster cannot always provide sufficient resources for ev- ery user. A computing cluster with cloud resources assisted could let users obtain computing resource faster while local machine are fully loaded. We collected traces from workstations in our department to understand how users use machines for computing and try to improve the scheduler. To make a sys- tem scheduler dispatch an incoming job to a local machine or to the cloud, we provide a task run time advisor for the scheduler making decisions accu- ractely. The run time advisor is a support vector regression model which is constructed by historical information. The prediction error is less than 20% error of mean run time in predicting CPU-bound open source projects. Our evaluation experiment is a data driven approach that the trace is collected from workstations in NTU CSIE.