RCP: Myrinet上的可信賴控制協定

Abstract

在叢集式(cluster)及分散式(distributed)電腦架構上，高速的傳輸速率 是非常重要的，以往大都使用乙太網路(Ethernet)做為傳輸的媒介，但由 於乙太網路的傳輸速率低、延遲時間長，使得系統整體效率變得很低。 Myrinet是一個具有gigabit-per-second 的高速網路，它可以使用在各種 不同的平台，但由於Myrinet網路介面卡上的軔體MCP (Myrinet Control Program)並不可信賴(reliable)，因此我們的重點著重在移植Myrinet到 OSF/1 AD3.0上，並修改MCP，使其在MCP layer就能夠保證可信賴，讓上 層的系統核心不需要顧慮可信賴的問題。我們的做法是在MCP中加上 buffer流量控制(flow control)、重送(retransmission)及錯誤訊息傳遞 來達成可信賴。實驗的結果不僅讓Myrinet達到可信賴，同時最高傳輸速 率可達563Mbit/s。 High-speed transmission is very important for cluster-based and distributed computing architectures. Previous researches focus on using the Ethernet as the communication media. Because of the low throughput and high latency of Ethernet, system throughput degrades greatly. Myrinet is a high-speed network, which can transmit data at gigabit per second. It can also be used on several different platforms. The firmware on Myrinet interface card, the Myrinet Control Program (MCP), is not reliable that motivates us to port the Myrinet device driver to the OSF/1 AD3.0 and modify the MCP to ensure system reliability. Our approaches are to add the buffer flow control, retransmission, and error message passing. These mechanisms guarantee reliable transmission at the MCP layer, as a result that the higher level of system kernel need not deal with the problem of reliability. Experimental results show that the modified Myrinet is reliable and the throughput is 563Mbits/s maximum.