Improve HTTP/TCP Performance over ATM Networks: New Schemes and Performance Comparisons' zyxw Chin-Dong Wang and Ruay-Shiung Chang Department of Information Management National Taiwan University of Science and Technology Email: rschang@cs.ntust.edu.tw Abstract zyxwvutsrq This paper proposes a "selective packet retransmission zyxwvutsrqp " scheme for improving H7TF'CP performance when transmitting through ATM networks. zyxwvutsr In selective packet retransmission, we utilize the property zyxwvut of humank perception tolerance for errors to detennine whether zyxwvutsrqp to retransmit a corrupted TCP segment or not. For lossable data, such as imagt.. when error occurs because of cell lost, it will not be retransmitted. The simulations show that, f o r the same buffer size and traffics load, seleclive packet retransmission results in higher throughput than PPD, EPD, and plain TCP over ATM. 1. Introduction when ATM is internetworking with the existing network infrastructure (predominantly TCPIIP), some special signaling, addressing and routing protocols are needed. There are two methcds for implementing IP on an ATh4 network. The first is to implement a MAC layer to provide an emulation of IEEE 802 networks over ATM. The second approach replaces the data-;link layer of the protocol stack with an ATM-aware layer. However, in either way certain features of ATM technology lead to substantial degradation of TCPDP connection throughput under congestion conditions. For example, as cell size is fixed, the last cell will carry a padding (wasted space) and the 5-bytes headers also increase the overhead. The poor performance of TCPDP over ATM has been observed in recent studies [ zyxwvuts 1,2] and a number of remedies has been proposed [2-61. Most of these mechanisms to improve TCPIIP performance over ATM networks require substantial complication of the ATM switches. In this paper, we propose a new method to improve the H'ITPRCP performance over ATM[ networks, called the selective packet retransmission scheme. The principle of this method is to categorize the transfer data into two types-lossable and unlossable. For lossable packets, they don't time out and retransmit when there are some errors. \Ne use the data transferred by HTTP (HyperText Transfer Protocol) to demonstrate the !scheme. Intuitively, since H'ITP data are now a major source of network traffic, when its performance improves, so will the performance of !:he whole network. Simulation results indeed corroborate with our expectations that performance can be improved with a litter sacrifice in TrP data (graphic or image data) precision. The remainder of this paper is organized as follows. The detail of our proposed mechanism is presented in :Section 2. Analysis and performance simulation results are illustrated in Section 3 and Section 4 concludes the paper. 2. Protocol Operations In this section we will describe the modifications needed for HTTP and TCP protocols to reduce the unnecessary retransmissions and improve packet latency. HTTF is a simple protocol. The client establishes a TCP connection to the server, issues a request, and reads back the server's response. The server denotes the end of its response by closing the connection. The client parses the HTML document according to tags and fetches the inlined files if necessary. For example, text files, image files, voice files, and execution files. According to the statistics in [71, GIF graphics and Text files are the most popular, each at about 2.5%. PostScript and PEG files are just over 1%. All other fo'rmats are below 1%. These files are segmented into TCP packets for transmission. When the underlying networks are Am, TCP packets are further sliced into ATM cells. If a cell is lost due to buffer overflow or other reasons in the ATM network, the whole packet containing this cell zyxw has to be retransmitted. This is a great. waste of bandwidth and the slow start congestion control [81 algorithm of TCP make,s it even worse. Therefore, it is appropriate to reconsider the I This research is supported in part by NSC under contract numbers NSC86-2622-E-007-OCIl and NSC86-2213- E-01 1-056. 0-8186-8227-2197 $10.00 zyxwvutsrqp 0 1997 IEEE 663