INTERNET DRAFT S. Bandyopadhyay draft-shyam-rt-pkt-transport-00.txt March 18, 2014 Intended status: Informational Expires: September 18, 2014 An approach of transportation of RT packets through IP switch based networks to achieve the best performance. draft-shyam-rt-pkt-transport-00.txt Abstract This document intends to find out the size of an IP packet at which VoIP applications will produce the best result. It emphasizes the physical phenomenon because of which ATM networks perform better than the IP switch based networks and tries to come out with an approach by which IP switch based network can perform as good as ATM network for the processing of real time traffic. Status of this Memo This Internet-Draft is submitted in full conformance with the provisions of BCP 78 and BCP 79. Internet-Drafts are working documents of the Internet Engineering Task Force (IETF). Note that other groups may also distribute working documents as Internet-Drafts. The list of current Internet- Drafts is at http://datatracker.ietf.org/drafts/current/. Internet-Drafts are draft documents valid for a maximum of six months and may be updated, replaced, or obsoleted by other documents at any time. It is inappropriate to use Internet-Drafts as reference material or to cite them other than as "work in progress." This Internet-Draft will expire on September 18, 2014. Copyright Notice Copyright (c) 2014 IETF Trust and the persons identified as the document authors. All rights reserved. This document is subject to BCP 78 and the IETF Trust's Legal Provisions Relating to IETF Documents (http://trustee.ietf.org/license-info) in effect on the date of publication of this document. Please review these documents carefully, as they describe your rights and restrictions with respect to this document. Bandyopadhyay Expires September 18, 2014 [Page 1] Internet Draft RT packet transport through IP switches March 18, 2014 1. Introduction Traditionally ATM network performs faster than the IP switch based network. The difference becomes more prominent for real time applications. Whereas they have disadvantages as far as bandwidth usages is concerned compared to the IP-switch based network. This document tries to address approaches for IP-switch based network to process real-time applications as fast as ATM network. 2. Processing of real time packets Here is an attempt to come out with a solution for IP switch based network to operate in the most user-friendly manner to transport data traffic (IP) as well as real time (RT) traffic (as RTP[1] packet) in the existing 32-bit system. In case of IP routing/switching entire packet gets collected at the intermediate router/switch and forwarded based on the forwarding table. Inside the switch/router the variable length IP packet gets fragmented into smaller size frames at the ingress side. The frames gets transported through the switching fabric with proper priority mechanism (to support QoS) and then reassembled at the egress side and passed through the media for the next hop. In case of ATM, packets get fragmented at the ingress edge devices into small size cells. Entire packet gets transported as a stream of cells and gets collected at the egress edge device. The success of ATM over IP routing as far as speed is concerned is due to the fact that the latency gets reduced as the entire packet does not get collected, fragmented and reassembled at the intermediate nodes. So, in case of IP switch based network, if RT packets can be passed without getting fragmented inside the switch, better performance can be expected. i.e. one RT packet needs to get to fit inside one internal frame of the switch fabric. Additionally, to make this approach successful, maximum size of MPLS[2] label stack has to be defined. Inside the switch all the IP packets will be assumed to carry same number of MPLS labels whether they are having one or the maximum in real sense. In fact, to reduce overhead, this limit should be the minimum number of labels needed to satisfy all sorts of features supported by MPLS. i.e. label stacking of depth n (without limit) needs modification. If minimum frame size is selected to fit one RTP packet, overhead becomes too high due to very large (40 bytes: 20 bytes IP + 8bytes UDP + 12 bytes RTP) packet header. Again, if large frame size is used, fragmentation loss becomes too high for the small size packets (say, 40 bytes IP packets). So, a compromise is needed that will give a better result based on the IP packet size distribution. Frame size Bandyopadhyay Expires September 18, 2014 [Page 2] Internet Draft RT packet transport through IP switches March 18, 2014 is selected based on the minimum value of the overhead due to the fragmentation loss of data packet as well as the overhead as header of the RT packets. Studies show that primarily IP data packets of three different sizes are found common in nature. Almost ~50% packets of size 40 bytes (TCP ACK), ~20% packets of size 576 bytes (path MTU set by X.25) and ~30% packets of size 1500 bytes (path MTU set by ethernet) Other packets are less compared to the above three categories and almost evenly distributed. For the sake of simplicity of calculation, traffic of the first three categories are only considered. Payload of the data traffic is the actual IP packet size where as the payload of RT traffic is the payload inside RTP packet. If totBytes are to be transported across the internet and dataPcnt be the %of data traffic, totBytes*dataPcnt/100 = data traffic and (100-dataPcnt)*totBytes/100 = RT traffic; Out of data traffic 50% of 40 bytes length; 20% of 576 bytes length;& 30% of 1500 bytes length. If totDataPkts be the total data packets, totDataPkts*(50*40/100 + 20*576/100 + 30*1500/100) = totBytes*dataPcnt/100; or, totDataPkts*58520 = totBytes*dataPcnt; Let totBytes = 58520*100, for the ease of calculation; i.e. totDataPkt = dataPcnt*100; 40 bytes packets = 50*totDataPkt/100 i.e. 50*dataPcnt 576 bytes packets = 20*totDataPkt/100 i.e. 20*dataPcnt 1500 bytes packets = 30*totDataPkt/100 i.e. 30*dataPcnt RT packets = totBytes * (100 - dataPcnt)/100 = 58520 * (100-dataPcnt); If n is considered to be the depth of MPLS label stack, inside the switch, actual size of 40 bytes packet = 40+4*n bytes, 576 bytes packet = 576+4*n bytes & 1500 bytes packet = 1500+4*n bytes Let frameSize be the payload of a frame (excluding the frame header) inside the switch. If a RT packet fits exactly inside frameSize, RT packet payload = (frameSize-40-4*n) bytes; Bandyopadhyay Expires September 18, 2014 [Page 3] Internet Draft RT packet transport through IP switches March 18, 2014 Total overhead = packet header overhead (of RT packets) + fragmentation overhead (of data packets); If a plot is drawn for frameSize = 40+4*n+1 to 1500+4*n for different dataPcnt (with dataPcnt=80 to 100) minimum of overhead are found at frameSize = (84, 101, 118, 126 and 152) for n==3; frameSize = (119, 127 and 152) for n==4 and at frameSize = (118, 127 and 152) for n==5. Actual data of the IP traffic has to be collected to get the best result. As dataPcnt increases minimum values are found at a lower frameSize and it gives better result with the higher range for lower dataPcnt. With average IP packet size 585 bytes, switches will encounter a loss of 4*(n-1) bytes for packets that will need only one label. In order to make this scheme work, a standard for maximum label stack size has to be defined. RTP packet size also has to be standardized. The same scheme is applicable to all the switching systems where IP packets get transported in hop by hop basis unlike the way it works in ATM networks. 2.1. Dual mode operation Inside ingress as well as in the egress card, packets need to follow certain functional steps. In order to maximize the output, a series of processing units work in pipeline mode for these operations. Ingress service cards need to act in dual mode to process RT packets and non-RT packets. i.e. the RT packets should follow a direct path that won't need fragmentation and related complexities before they are sent to the VOQs (virtual output queues, where from packets gets picked up to be sent to the switching fabric). Whereas other packets need to follow a different path for fragmentation operations. This will prevent a RT packet to be blocked by the fragmentation procedure of not-RT packets that arrive in the service card prior to the arrival of RT packet. So, mere mapping of RT packet size with the frameSize of switch fabric will not achieve the speed of ATM switches. Simulation studies show that significant improvement is achieved once RT packets are directly sent to VOQs after the operation of label processing. It will be worth to study by the hardware people to figure out whether entire set of data can be placed into queues based on their priorities and segmentation operation is done in each queue in parallel mode before putting the frames into their respective VOQs. Entire operation will be lot costlier, but simulation result shows that in such case, RT packets need not be restricted to fixed size cells. Standardization of label stack depth need not be imposed as well. Bandyopadhyay Expires September 18, 2014 [Page 4] Internet Draft RT packet transport through IP switches March 18, 2014 2.2. Expected changes at the application layer IP packets with size 576 in most of the cases come out of those TCP layers that do not process maximum path-MTU and takes the default one that was set during X.25. The 576 factor can be corrected very easily with path-MTU set to 1500. Also, in case of next generation IP, header of IP packets will be different. Switch fabric frame size needs to be determined keeping these two factors in mind. With the existing 32-bit system, frame size (excluding the frame header) of 152 and 127 are most viable solution in general for label stack depth=3,4 &5. 3. Security Consideration This document does not include any security related issues. 4. Acknowledgments The author would like to thank to Professor Amitava Datta of University of Western Australia for his review and constructive comments. 5. Normative References [1] Schulzrinne, H., Casner, S., Frederick, R., and V. Jacobson. "RTP: A Transport Protocol for Real-Time Applications", RFC 3550, July 2003. [2] Rosen, E., Viswanathan, A. and R. Callon, "Multiprotocol Label Switching Architecture", RFC 3031, January 2001. 6. Author's Address Shyamaprasad Bandyopadhyay HL No 205/157/7, Inda Kharagpur 721305, India Phone: +91 3222 225137 e-mail: shyamb66@gmail.com Bandyopadhyay Expires September 18, 2014 [Page 5]