篇名 | Performance Analysis of Direct Access to Remote Memory |
---|---|
卷期 | 166 |
作者 | Peng-Kai Hsu 、 Tai-Hua Hsiao 、 Wen-Shuo Chou 、 Hung-Yi Teng |
頁次 | 076-081 |
關鍵字 | Rack Scale Architecture 、 Ethernet 、 PCIe 、 Remote Memory Access 、 DMA |
出刊日期 | 201606 |
Rack Scale Architecture (RSA) is one of common infrastructure of cloud OS or NFV system. In RSA, the interconnection technology has a direct bearing on its performance. So far, there are two major interconnection technologies should be concerned: Ethernet and PCI Express (PCIe). Although Ethernet dominates current computer networking, it seems to become a performance bottleneck due to several existed drawbacks. On the other hand, PCIe is a high-speed parallel bus which is originally designed to connect I/O peripherals to a CPU. However, PCIe can be extended into an intra-rack inter-server network and become a promising solution to resolve the drawbacks of Ethernet. Therefore, in this paper, we implemented a prototype of a PCIe-based host-to-host architecture, conducted experiments of remote memory access (RMA) and compared with an Ethernet-based architecture. Experimental results showed that when message size is 10,000 bytes, the write throughput of the PCIe architecture can reach around 19 Gb/s which is much higher than that of the Ethernet-based architecture (around 9Gb/s). But, we found that the read throughput and the write/read latency of the PCIe architecture are worse than that of the Ethernet-based architecture. It is probably because (1) the DMA engine does not be optimized and (2) each read operation needs to go through 16 PCIe bridges and 2 NTBs in our experiments. These results could provide a better understanding of the PCIe-based RSA technologies.