H3C Key words: QoS, quality of service Abstract: The Ethernet technology is widely applied currently. At present, Ethernet is the leading technology in various independent local area networks (LANs), and many Ethernet LANs have been part of the Internet. With the development of the Ethernet technology, most common Internet users access the Internet through Ethernet. To implement end-to-end QoS throughout the network, you must guarantee QoS for Ethernet. To do this, Ethernet switching devices must use the QoS technology to provide different QoS guarantees for different types of traffic flows, especially those traffic flows with higher demand for delay and jitter guarantees. Acronyms: Acronym Full spelling QoS Quality of Service Copyright 2007 Hangzhou H3C Technologies Co., Ltd. Page 1/15
Table of Contents 1 Overview...3 2 Basic Networking Structure...3 3 Features...4 3.1 Service Model...4 3.2 Traffic Classification...4 3.3 Traffic Policing...5 3.4 Priority Marking...6 3.5 Queue Scheduling...8 3.6 Congestion Avoidance...10 3.7 Traffic Shaping...12 3.8 Policy Routing...13 4 QoS Processing Procedure on the S9500 Series...14 Copyright 2007 Hangzhou H3C Technologies Co., Ltd. Page 2/15
1 Overview On traditional packet switching networks, switches and routers treat all packets equally and handle them using the first in first out (FIFO) policy. This service is called best-effort. It delivers packets to their destinations as possibly as it can, without any guarantee for delay and jitter. With the development of computer networks, more and more traffic such as voice, video, and critical data which is sensitive to bandwidth, delay, and jitter is transmitted over networks. This enriches the services resources on a network greatly. On the other hand, there is a higher demand for the Quality of Service (QoS) of network transmission. The Ethernet technology is widely applied currently. At present, Ethernet is the leading technology in various dependent local area networks (LANs), and many Ethernet LANs have been part of the Internet. With the development of the Ethernet technology, most common Internet users access the Internet through Ethernet. To implement end-to-end QoS throughout the network, you must guarantee QoS for Ethernet. To do this, Ethernet switching devices must use the QoS technology to provide different QoS guarantees for different types of traffic flows, especially those traffic flows with higher demand for delay and jitter guarantees. 2 Basic Networking Structure Figure 1 Basic networking structure Copyright 2007 Hangzhou H3C Technologies Co., Ltd. Page 3/15
3 Features 3.1 Service Model A service model refers to a set of end-to-end QoS functions. The simplest service model is the Best-Effort model adopting the FIFO policy. It delivers packets to their destinations as possibly as it can, without any guarantee for delay and jitter. The Diff- Serv model was introduced to implement QoS for network transmission. The Diff-Serv model is a multi-service model. It provides QoS services for each packet according to the QoS parameters specified for the packet, thus satisfying differentiated QoS demands. The Diff-Serv model is used to implement end-to-end QoS for some critical services. The S9500 series support the Diff-Serv model. 3.2 Traffic Classification To specify different QoS parameters for packets of different levels, the Diff-Serv model must classify the network traffic first. Traffic classification organizes packets with different characteristics into different classes using classification rules. A classification rule is a filter rule configured to meet your management requirements. It can be very simple. For example, you can use a classification rule to identify traffic with different priorities according to the ToS field in the IP packet header. It can be very complicated too. For example, you can use a classification rule to identify the packets according to the combination of link layer (Layer 2), network layer (Layer 3), and transport layer (Layer 4) information including MAC addresses, IP protocol, source addresses, destination addresses, port numbers of applications, and so on. Generally, the traffic classification criterion is limited in the header of an encapsulated packet. Contents of packets are rarely adopted for traffic classification. The S9500 series support Layer 2, Layer 3, and Layer 4 ACL rules for traffic classification. Such ACL rules can classify packets based on source MAC addresses, destination MAC addresses, VLAN IDs, source IP addresses, destination IP addresses, source TCP/UDP port numbers, destination TCP/UDP port numbers, protocol types, IP precedence, ToS precedence, DSCP precedence, and whether packets are fragmented. Copyright 2007 Hangzhou H3C Technologies Co., Ltd. Page 4/15
3.3 Traffic Policing To use limited network resources to provide customers with better services, you can enable traffic policing on the incoming port for the traffic of the specified customers, thus making the traffic adapt to the network resources assigned to it. Traffic policing uses token buckets for traffic control. Figure 2 Traffic policing Figure 2 depicts the processing procedure of traffic policing. First, packets are classified and the packets with the specified characteristics enter the token bucket for processing. If the token bucket has enough tokens for sending the packets, the packets can pass through; otherwise, the packets are dropped. In this way, you can control the traffic of a certain class of packets. The system puts tokens into the bucket at the set rate. You can set the capacity of the token bucket. When the token bucket is full, the extra tokens will overflow and the number of tokens in the bucket stops increasing. When the token bucket processes packets, if it has enough tokens for sending these packets, the packets are sent, and at the same time, the corresponding number of tokens are taken out of the bucket. If the token bucket does not have enough tokens for sending these packets, these packets are dropped. Therefore, the traffic rate is restricted under the rate of generating tokens, thus implementing traffic control. Copyright 2007 Hangzhou H3C Technologies Co., Ltd. Page 5/15
The S9500 series support traffic policing with the granularity of 8 kbps. 3.4 Priority Marking Through marking different priorities for packets, you can identify the service levels of different packets. The S9500 series can perform priority marking for specific packets. ToS precedence, differentiated services codepoint (DSCP) precedence, and 802.1p precedence can be marked. These priority types apply to different QoS models and are defined in different models. The following part introduces IP precedence, ToS precedence, DSCP precedence, 802.1p precedence, and EXP precedence. I. IP precedence, ToS precedence, and DSCP precedence Figure 3 IP precedence, ToS precedence, and DSCP precedence As shown in Figure 3, the ToS field of the IP header contains 8 bits: the first three bits (0 to 2) represent IP precedence from 0 to 7; the following 4 bits (3 to 6) represent a ToS value from 0 to 15. In RFC2474, the ToS field of the IP header is redefined as the DS field, where a DiffServ code point (DSCP) precedence is represented by the first 6 bits (0 to 5) and is in the range 0 to 63. The remaining 2 bits (6 and 7) are reserved. II. 802.1p precedence 802.1p precedence lies in Layer 2 packet headers and is applicable to occasions where the Layer 3 packet header does not need analysis but QoS must be guaranteed at Layer 2. Copyright 2007 Hangzhou H3C Technologies Co., Ltd. Page 6/15
Figure 4 802.1Q Ethernet frame format As shown in the figure above, each host supporting the 802.1Q protocol adds a 4- byte 802.1Q tag header after the source address of the former Ethernet frame header when sending the packet. The 4-byte 802.1Q tag header contains a 2-byte Tag Protocol Identifier (TPID) whose value is 8100 and a 2-byte Tag Control Information (TCI). TPID is a new class field defined by IEEE to indicate that the current packet is 802.1Q-tagged. Figure 5 describes the detailed contents of an 802.1Q tag header. Figure 5 802.1p precedence In the figure above, the 3-bit priority field in the TCI field is 802.1p priority in the range of 0 to 7. The three bits specify the precedence of the frame. Eight precedence values are used to determine which packets are sent preferentially when congestion occurs. The precedence is called 802.1p precedence because applications related to the precedence are defined in detail in the 802.1p specifications. To provide differentiated services for VLAN VPN or QinQ frames, you must classify frames by VLANs or 802.1p precedence in their inner VLAN tags. The inner VLAN and 802.1p precedence of a packet determines its queue scheduling priority and drop precedence. The 802.1p precedence of the inner VLAN tag of a packet determines the scheduling priority and drop precedence of a packet at the egress. Figure 6 802.1p precedence mapping Copyright 2007 Hangzhou H3C Technologies Co., Ltd. Page 7/15
III. EXP precedence Figure 7 MPLS label In an Ethernet MPLS packet, there is a shim between the Layer 2 header and Layer 3 data. You can use the reserved fields in the shim, a 3-bit EXP to determine the scheduling priority and drop precedence of the packet. You can classify MPLS packets by their EXP precedence and determine the scheduling priority and drop precedence of MPLS packets at the egress. You can map the DSCP precedence of IP packets to the EXP precedence and use the EXP precedence to determine the scheduling priority and drop precedence of MPLS packets at the egress. Figure 8 EXP precedence marking 3.5 Queue Scheduling When the network is congested, the problem that many packets compete for resources must be solved, usually through queue scheduling. The S9500 series support two queue scheduling algorithms: strict priority (SP), and weighted round robin (WRR). Copyright 2007 Hangzhou H3C Technologies Co., Ltd. Page 8/15
I. SP queue scheduling algorithm High priority Queue 7 Packets sent via this interface Queue 6 Packets sent Queue 5~2 Classify dequeue Sending queue Queue 1 Low priority Queue 0 Figure 9 Diagram for SP queueing SP queue scheduling algorithm is dedicated to critical service applications. The key feature of mission-critical applications is that they require preferential service to reduce the response delay when congestion occurs. Assume that there are eight output queues on a port and the SP queueing classifies the eight output queues on the port into eight classes, which are queue 7, queue 6, queue 5, queue 4, queue 3, queue 2, queue 1, and queue 0 in the descending order of priority. SP schedules the packets in a strict priority order. It sends the packets in the queue of the highest priority first, and sends packets in a queue of a lower priority only when the queue of a higher priority is empty. You can put critical service packets into the queues with higher priority and put non-critical service (such as e-mail) packets into the queues with lower priority. In this case, critical service packets are sent preferentially and non-critical service packets are sent when critical service groups are not sent. The SP mechanism has its disadvantage. When congestion occurs and if high-priority queues are occupied for a long time, the packets in the lower-priority queues are starved before obtaining services. Copyright 2007 Hangzhou H3C Technologies Co., Ltd. Page 9/15
II. WRR queue scheduling algorithm A switch port supports eight output queues. WRR queue-scheduling algorithm schedules all the queues in turn and every queue can be assured of a certain service time. Assume there are eight priority queues on a port. WRR configures a weight value for each queue, which is w7, w6, w5, w4, w3, w2, w1, and w0. The weight value indicates the proportion of obtaining bandwidth. On a 100 M port, configure the weight value of WRR queue-scheduling algorithm as 50, 30, 10, 10, 50, 30, 10, and 10 (corresponding to w7, w6, w5, w4, w3, w2, w1, and w0 in order). In this way, the queue with the lowest priority can get 5 Mbps bandwidth at least, thus avoiding the disadvantage of SP queue-scheduling that the packets in queues with lower priority may not get service for a long time. Another advantage of WRR queuing is that: though the queues are scheduled in order, the service time for each queue is not fixed; that is to say, if a queue is empty, the next queue will be scheduled. In this way, the bandwidth resources are made full use. 3.6 Congestion Avoidance When the network is congested, common network devices adopt tail drop to avoid congestion. That is, when the queue length reaches the upper threshold, all the newly arriving packets are dropped. However, if plenty of TCP traffic is dropped, which will cause TCP timeout, the slow start and congestion avoidance mechanisms of TCP will be triggered, thus reducing TCP traffic. If a queue drops packets of multiple TCP sessions at the same time, slow start and congestion avoidance mechanisms will be triggered for these TCP sessions at the same time. This is called global TCP synchronization. In this case, these TCP sessions reduce the size of traffic sent to the queue at the same time, so that the traffic sent to the queue is less than the bandwidth of the queue, thus reducing the utilization of the line. On the other hand, the size of the traffic sent to the queue is not stable but fluctuates between the maximum bandwidth and a very small traffic size. The S9500 series adopt the Weighted Random Early Detection (WRED) mechanism to avoid global TCP synchronization. You can set the upper threshold and lower threshold for a queue. When the queue length is smaller than the lower threshold, no Copyright 2007 Hangzhou H3C Technologies Co., Ltd. Page 10/15
packet is dropped; when the queue length is between the lower threshold and the lower threshold, WRED begins to drop packets randomly, and the drop probability increases as the queue length increases; when the queue length is bigger than the upper threshold, all newly arriving packets are dropped. WRED drops packets randomly, thus avoiding global TCP synchronization. When the sending rate of a TCP session slows down after its packets are dropped, the other TCP sessions remain in high packet sending rates. In this way, some TCP sessions remain in high packet sending rates in any case, and the link bandwidth can be fully utilized. If the current queue length is compared with the upper threshold and lower threshold to determine the drop policy, bursty traffic is not fairly treated and proper data transmission is affected. To solve this problem, WRED compares the average queue size with the lower threshold and upper threshold to determine the drop policy. The average queue size reflects the queue size change trend but is not sensitive to bursty queue size changes, and thus bursty traffic can be fairly treated. On a S9500 switch, you can set the exponential factor for average queue length calculation, upper threshold, lower threshold, and drop probability for packets with different precedence values respectively to provide differentiated drop policies. When congestion occurs, the S9500 switch drops packets as soon as possible to release queue resources and try not to assign packets to high-delay queues in order to eliminate congestion. A S9500 switch can assign drop levels to packets according to their 802.1p precedence, that is, color the packets, or assign drop levels through priority marking. The drop level can be 0, 1, or 2, which represent green, yellow, and red respectively. When congestion occurs, red packets are the first to be dropped, while green packets are the last to be dropped. You can set congestion avoidance parameters and thresholds for each queue and each drop level. The S9500 series support two drop algorithms: Tail drop: when packets are dropped, the drop policy for packets in a color (red, yellow, or green, assigned according to drop levels) is determined by the Copyright 2007 Hangzhou H3C Technologies Co., Ltd. Page 11/15
threshold set for the color. When the size of packets in a color (red, yellow, or green) exceeds the corresponding upper threshold, the system beings to drop newly arriving packets in this color. WRED drop algorithm: the drop levels are taken into account when packets are dropped by queue. When the size of packets in a color (red, green, or yellow) exceeds the lower threshold set for the color, the system begins to drop the packets in the color between the upper threshold and lower threshold according to a certain slope. When the size of packets in a color exceeds the upper threshold set for the color, the system begins to drop all packets in the color exceeding the upper threshold. 3.7 Traffic Shaping Traffic shaping controls the rate of output traffic, so that the traffic can be sent out at an even rate. Normally, traffic shaping is applied on a device to adapt its output rate to the input rate of its connected downstream device so as to avoid unnecessary packet drop and congestion. It differs from traffic policing mainly in that traffic shaping buffers packets exceeding the rate limit so that packets are sent out at an even rate, while traffic policing drops packets exceeding the rate limit. However, traffic shaping introduces additional delay while traffic policing does not. The S9500 series support port-based traffic shaping, that is, traffic shaping can be implemented to all traffic on a port. It also supports queue-based traffic shaping on a port. Copyright 2007 Hangzhou H3C Technologies Co., Ltd. Page 12/15
3.8 Policy Routing Figure 10 Policy routing application scenario The S9500 series can classify packets first and then configure traffic redirecting for a certain class of packets to implement policy routing. As shown in Figure 10, the S9500 switch first classifies packets based on source IP addresses and destination IP addresses to identify packets whose source IP addresses are private address while whose destination IP address are public addresses. Then you can use policy routing to redirect such packets to the NAT device for address translation and then to the Internet. Copyright 2007 Hangzhou H3C Technologies Co., Ltd. Page 13/15
4 QoS Processing Procedure on the S9500 Series Figure 11 QoS processing procedure on the S9500 series The S9500 series use traffic classification to classify traffic based on source MAC addresses, destination MAC addresses, Ethernet types, VLANs, 802.1p priority, IP protocol, source IP addresses, destination IP addresses, application port numbers, ICMP packet types, IP precedence, ToS, DSCP, EXP, and VLAN IDs and 802.p priorities in the inner VLAN tags of QinQ frames. After classifying traffic into different classes, besides simply permitting a class of packets to pass through or dropping a class of packets, the S9500 series provide a policy control list (PCL) to perform the following actions for the traffic flows: traffic policing, traffic accounting, marking QoS parameters (including 802.1p priority, DSCP, EXP, and drop precedence), traffic mirroring, traffic redirecting, and specifying the output queue. After packets are marked with different drop levels through priority mapping, the congestion avoidance module determines the drop policies for packets based on the user-defined drop mode and the upper threshold and lower threshold set for each color. With tail drop adopted, when the size of packets in a color (red, yellow, or green) exceeds the upper threshold set for the color, the system begins to drop newly arriving packets in the color. With WRED drop mode adopted, when the size of packets in a color (red, green, or yellow) exceeds the lower threshold set for the color, Copyright 2007 Hangzhou H3C Technologies Co., Ltd. Page 14/15
the system begins to drop the packets between the upper threshold and lower threshold according to a certain slope. When the size of packets in a color exceeds the upper threshold set for the color, the system begins to drop all packets in the color exceeding the upper threshold. After congestion avoidance is completed, the packets permitted to be forwarded are assigned to the corresponding queues. The queue scheduling module uses SP or WRR queue scheduling algorithm to schedule packets. When forwarding packets, the output port performs traffic shaping for outbound traffic based on the token bucket size. Copyright 2007 Hangzhou H3C Technologies Co., Ltd. All rights reserved. No part of this manual may be reproduced or transmitted in any form or by any means without prior written consent of Hangzhou H3C Technologies Co., Ltd. The information in this document is subject to change without notice. Copyright 2007 Hangzhou H3C Technologies Co., Ltd. Page 15/15