- 會議論文
- OpenAccess
NCS 2019 Implement MPI Cloud Computing Platform with Shared Memory Architecture
摘要
在2016年AlphaGo成為了世界上第一套在圍棋上擊敗了職業九段棋士的人工智慧，在全世界掀起了機器學習的風潮。對神經網路演算法來說，背後必定不可少的是非常大量的訓練資料庫，能夠有效、快速地運算大量資料顯得十分重要。目前大多使用雲端運算與多核心平行運算等方法作為處理大量運算資料的解決方法。我們提出結合了雲端運算與多核心平行運算，實現了既可最大化使用運算資源，又可以藉由多機連線擴充運算資源的多核心雲端運算平臺：Message Passing Interface Cloud Computing Platform with Shared Memory Architecture（MPICSM）。傳統MPI運算平臺以使用TCP/IP Protocol作為訊息溝通平臺為原則，而我們建構多核心機器的本機溝通平台架構，透過提供建立多進程API的方式，以系統核心多執行緒的概念，建立多個進程以達到多核心平行運算，並且使用共享記憶體架構取代網路連線作為同一運算機器內部進程的訊息傳遞介面。MPICSM提供了符合MPI Standard 3.1的API，讓使用者在使用MPICSM編寫平行運算程式時能輕易上手。我們實測驗證MPICSM的運算效能驗證結果顯示，使用共享記憶體架構作為溝通手段時，資料傳輸效能較網路傳輸提升約6倍，在溝通密集的情況下，增加使用共享記憶體傳遞訊息的比例，能提升整體運算效能。
並列摘要
Since AlphaGo became the world's first artificial intelligence to defeat the professional nine-segment chess player in Go. Machine learning is now all the rage around the world. In the case of the neural network, it is indispensable to have a very large number of training databases. Therefore, processing mass data efficiently and quickly becomes a significant issue. For mass data processing, cloud computing and multi-core computing are considered popular solutions. We merge the cloud computing and the multi-core computing to realize a multi-core cloud computing platform that can not only maximize the utility of computing resources but expand computing resources through connecting the multi-computer : Message Passing Interface Cloud Computing Platform with Shared Memory Architecture(MPICSM). To realize cloud computing, we use the TCP/IP Protocol to connect computers. In addition to this, by process building API, several computing processes are built to realize multi-core computing. Shared memory architecture plays the role of connecting the processes in a single computer. MPICSM provides APIs that comply with MPI Standard 3.1, users can quickly create parallel processing programs with it. To verify the feasibility of MPICSM in parallel processing, we use the Leibniz pi, K-Means Clustering, and AES algorithm to check the performance of MPICSM. The verification results show that when using the shared memory architecture as a means of communication, the data transmission performance is about 6 times higher than the network transmission. With the result above, several programming recommendations using the MPICSM platform are proposed.