IEEE 802.17 Resilient Packet Ring Background and Overview
Fredrik Davik, Mete Yilmaz, Stein Gjessing, Necdet Uzun
Abstract — The IEEE Working group P802.17 is standardizing a new ring topology network architecture, called the Resilient Packet Ring (RPR), to be ud mainly in metropolitan and wide area networks. This paper prents a technology background, gives an overview, and explains many of the design choices the RPR working group faced during the development of the standard. Some major architectural features are illustrated and compared by showing performance evaluation results using the RPR simulator developed at Simula Rearch Laboratory using the OPNET Modeler simulation environment.
手抄报内容Index Terms— Communications, Networking, MAN, WAN, Ring networks, Spatial reu, Fairness.
Fredrik Davik is with Simula Rearch Laboratory and Ericsson Rearch
Mete Yilmaz is with Cisco Systems
Stein Gjessing is with Simula Rearch Laboratory and is a visiting scholar at Department of Computer Engineering, San Jo State University
Necdet Uzun is with Cisco Systems
A shorter version of this report is to appear in IEEE Communications Magazine March 2004. Simula Rearch Laboratory, Technical Report 11-2003, December 2003
I.I NTRODUCTION
T
he Resilient Packet Ring (RPR, IEEE 802.17) is the latest development in a ries of ring bad network protocols standardized by IEEE [6]. IEEE 802.5 Token Ring [7] and IEEE 1596 Scalable Coherent Interface (SCI) [8] are examples of other ring bad IEEE standards. Packet ring bad data networks were pioneered by the Cambridge Ring [10], followed by other important network architectures, notably MetaRing [2], FDDI [12], ATMR [13] and CRMA-II [14].
Rings are in general built using veral point-to-point connections. When the connections between the stations are bidirectional, rings allow for resilience (a frame can reach its destination even in the prence of a link failure). A ring is also simpler to operate and administrate than a complex mesh or an irregular network.
淘宝网品牌女装
Networks deployed by rvice providers in the MANs or WANs are often bad on SONET/SDH rings. Many SONET rings consist of a dual-ring configuration in which one of the rings is ud as the back-up ring that remains unud during normal operation and utilized only in the ca of failure of the primary ring [1]. The static bandwidth allocation and network monitoring requirements increa the total cost of a SONET network. While Gigabit Ethernet does not require static allocation and provides cost advantages; it cannot provide desired features such as fairness and fast (<50ms) auto-restoration.
In order to provide efficient, carrier class packet transport, some companies started to develop proprietary ring technologies. For example, Cisco Systems developed the Dynamic Packet Transport (DPT) [20] technology bad on the Spatial Reu Protocol (SRP) [19], and Nortel Networks developed the OPTera Packet Edge technology [16].
In order to standardize the new initiatives, IEEE was approached. One of the goals of the new standard is to utilize the simplicity of ring networks and u the bandwidth of the dual-ring as efficiently as possible for high-speed data transmission in MANs and in WANs. Important goals also includes distribution of bandwidth fairly to all active stations while providing fast auto restoration. For a rapid and widespread deployment, the reu of existing physical layers is another important goal.
To achieve all of this, the IEEE working group P802.17 was formally started in March 2001 under the name Resilient Packet Ring. Since RPR is being standardized in the IEEE 802 LAN/MAN (Ethernet) families of network protocols, it can inherently bridge to other IEEE 802 networks and mimic a broadcast medium. RPR implements a Medium Access Control (MAC) protocol, for access to the shared ring communication medium, which has a client interface similar to that of Ethernet’s.
The rest of this paper is organized as follows: In ction II and III respectively ring network basics and RPR station design is discusd. The so-called fairness algorithm is the topic of ction IV, while ctions V, VI and VII treat topology discovery, resilience and bridging. Finally frame formats are outlined in ction VIII, and a conclusion is given. In order to demonstrate different operational modes, some performance figures are included and discusd. The scenarios have been executed on the RPR simulator model developed at Simula Rearch Laboratory and implemented in OPNET Modeler [15], according to the latest RPR draft standard as of November 2003 (v3.0).
II.R ING N ETWORK B ASICS
To facilitate discussion of design choices that were made in the development pha of RPR, this ction introduces some basic ring networking principles, not all implemented in RPR.
In unicast addressing (broadcast will be covered later), frames are added on to the ring by a nder station, that also decides on which of the two counter rotating rings (called ringlet 0 and ringlet 1 in RPR) the frame should travel to the receiving station. The destination address in the frame header might not be the exact address of the receiving station itlf. However, the station should, bad on the destination address, recognize that it is the receiving station for this frame. In this way the transmission on the ring (from nder to receiver) might be only one hop in a multi hop transmission from source to destination.
If a station does not recognize the destination address in the frame header, it transits the frame, i.e. the frame is forwarded to the next station on the ring. In RPR, the transit methods supported are cut through (the station starts to forward the frame before it is completely received) and store and forward.
To prevent frames, with a destination address recognized by no stations on the ring, to circulate forever, a time to live field (TTL) is decremented by all stations (as in RPR) or by one station (as in SCI) on the ring. Frames received with a TTL value of 0 is not pasd on to downstream stations (is stripped from the ring).
Figure 1. Destination Stripping and Spatial Reu illustrated on the outer ringlet. When a station recognizes that it is the receiver of a frame, it may copy the contents of the frame and let the frame traver the ring back to the nder (like in the Token Ring), it may nd back only an acknowledgement (if the station is able to receive the frame) or a
打印机什么牌子好
negative acknowledgement (if the station is unable to receive the frame) back to the nder (like in SCI), or it may remove the frame completely from the ring (like in RPR). When the receiving station removes the frame from the ring, the bandwidth otherwi consumed by this frame on the path back to the source, is available for u by other nding stations. This is generally known as spatial reu.
Figure 1 shows an example scenario where spatial reu is obtained on the outer ringlet; station 2 is transmitting to station 4 at the same time as station 6 is transmitting to station 9. Destination stripping with spatial reu was previously exploited in systems like MetaRing [2], ATMR [13], CRMA-II [14] and SCI [8].
The ring access method is an important design choice. A token may circulate the ring, so that the station holding the token is the only station allowed to nd (like in Token Ring). An alternative access method, called a “buffer inrtion” ring, was developed as early as 1974 [5][17]. Every station
on the ring has a buffer called an “inrtion buffer” (called a “transit queue” in RPR, e Figure 2) in which frames transiting the station may be temporarily queued. The station must act according to three simple rules. The first principle is that, the station will not add packets to the ring as long as there are packets in the inrtion buffer or packets in transit. Secondly, when there is no frame in transit, the station itlf is allowed to add a frame. Thirdly, if a passing frame arrives at a station when it has started to add a frame, the frame in transit is temporarily (for as long as it takes to complete the nding of the added frame) queued in the inrtion buffer. Obviously the three simple principles need some improvement to make up a full, working protocol that distributes bandwidth fairly. This has been studied before [3][11][4][9], and how this is achieved in RPR will be revealed in ction IV when the RPR fairness algorithm is discusd.
汗蒸减肥
Figure 2. The “inrtion buffer” or “transit queue” stores frames in transit, while the station itlf adds a frame (here a stations attachment to only one ring is shown).
III.S TATION DESIGN AND PACKET PRIORITY
The stations on the RPR ring implements a medium access control (MAC) protocol that regulates the stations access to the ring communication medium. All links around the ring is mandated to have the same capacity (called the full rate). The P802.17 working group has defined veral physical layer interfaces (reconciliation sublayers) for Ethernet (called PacketPHYs) and SONET/SDH [6]. The MAC also implements access points that clients can call in order to nd and receive frames and status information.
The RPR working group decided to implement a three level, class bad, traffic priority scheme. The objectives of the class bad scheme is to let class A be a low latency, low jitter class, class B be a class with predictable latency and jitter, and finally class C be a best effort transport class. It is worth to note that the RPR ring does not discard frames to resolve congestion. Hence when a frame has been added onto the ring, even if it is a class C frame, it will eventually arrive at its destination. The design decision behind this choice will be explained later
Class A traffic is divided into class A0 and A1, and class B traffic is divided into class B-CIR (Committed Information Rate) and B-EIR (Excess Information Rate). The two traffic class C and B-EIR are called Fairness Eligible (FE), for reasons that will become clear in ction IV.
The bandwidth around the ring is pre-allocated in two ways. The first is called "rerved" and can only be ud by class A0 traffic, and is equally rerved all around the ringlet. If stations are not using their pre-allocated A0 bandwidth, this bandwidth is wasted. In this way TDM-like traffic can be nt by RPR stations as A0 frames.
The other pre-allocated bandwidth is called “reclaimable”. A station that has class A1 or B-CIR traffic to nd, pre-allocates “reclaimable” bandwidth for the types of traffic. If not in u, such bandwidth can be ud by FE traffic. In addition, any bandwidth not pre-allocated is also ud to nd FE traffic. The distribution and u of unallocated and unud reclaimable bandwidth (FE bandwidth) is dynamically controlled by the fairness algorithm.圆的认识评课稿
衣衫褴褛A station’s rervation of class A0 bandwidth is broadcasted on the ring using topology messages (topology messages will be discusd later). Having received such topology messages from all other stations on the ring, every station calculates how much bandwidth to rerve for class A0 traffic. The
草原美酒
remaining bandwidth, called “unrerved rate” can be ud for all other traffic class. This in general means that the unrerved rate is the full rate minus the sum of all station’s A0 rervations. The standard claims that spatial reu of A0 bandwidth is possible but does not specify how this can be obtained.
An RPR station implements veral traffic shapers that limit and smooth the add- and transit traffic. There is one shaper for each of the traffic class A0, A1, B-CIR as well as one for FE traffic and one for non-A0 traffic. The shapers for class A0, A1, B-CIR and non-A0 traffic are preconfigured, while the FE shaper is dynamically adjusted by the fairness algorithm. The non-A0 shaper (also called the Downstream shaper) limits the amount of non class A0 traffic transmitted, ensuring that rerved bandwidth is available all around the ring. The rate of this shaper is the full rate minus the bandwidth rerved for class A0 traffic (i.e. unrerved rate).
When a station tries to nd more class B traffic than the B-CIR shaper allows, the rest of the class B traffic is nt as class B-EIR. Class B-EIR traffic has higher add-priority than class C traffic.
It is important that the preconfigured shapers are correctly t, in particular that the sum of the bandwidth allocated for A0, A1 and B-CIR traffic does not surpass the maximum bandwidth (the full rate).
It is not obvious that class A0 pre-allocated and rerved bandwidth may be spatially reud. For example stations 2 and 6 in figure 1 may both nd 10Mbit/c class A0 traffic to respectively stations 4 and 9. If this is the only class A0 traffic rerved, then ideally, the spatial reu feature should allow us to rerve only 10Mbit/s all around the ring. However, when the constant “rerved bandwidth” is calculated, it is calculated as the sum of all A0 allocations. If the downstream shaper is t to unrerved rate (full rate minus rerved rate), no spatial reu will be achieved for A0 traffic. If, on the other hand, the downstream shapers are t lower than the unrerved rate (becau of known spatial reu of A0 traffic), spatial reu of A0 traffic is indeed achieved. RPR does not contain mechanisms to ensure the intended spatial reu, and it may even not be mandated by the final standard to t the downstream shaper to any other value than the unrerved rate. Anyhow, if stations 2 and 6 can not be trusted to nd to non-overlapping gments of the ring, a total of 20 Mbit/c must be rerved around the ring for class A0 traffic. Misconfiguration of the downstream shaper may cau rious problems at run time.
Also class A1 and C-BIR traffic may be spatially reud, so that the total pre-allocated bandwidth on any link may be calculated taking spatial reu into consideration, in the same way as explained above for class A0 traffic. But also for the class of rvice, it is outside the scope of the RPR sta
ndard to enforce spatial reu. Hence, also here it might be wi to assume that all stations nd all their class A1 and B-CIR traffic all around the ring, and that the total pre allocated class A1 and B-CIR bandwidth is the sum of all station’s allocations. The difference between allocation of A1 and B-CIR bandwidth with or without spatial reu, only affects the calculation that is needed to ensure that the total ring capacity is not surpasd. Since unud class A1 and B-CIR bandwidth is reclaimable, this unud bandwidth may anyhow be ud by the fairness algorithm to nd FE-traffic (class B-EIR and class C traffic).
The minimum transit queue size is the maximum transfer unit that a station itlf may add (becau this is the maximum buffer size needed by the frames in transit while the station adds a new frame). Some flexibility for scheduling of frames from the add- and transit-path can be obtained by increasing the size of the transit queue. For example, a station may add a frame even if the transit queue is not completely empty. Also a larger queue may store lower priority transit frames while the station is adding high priority frames. The transit queue could have been specified as a priority queue, where frames with the highest priority are dequeued first. This was considered too complex and instead the working group decided that a station optionally may have two transit queues. Then high priority transit frames (class A) are queued in the Primary Transit Queue (PTQ), while class B a
nd C frames are queued in the Secondary Transit Queue (STQ). Forwarding from the PTQ has priority over the STQ and most types of add traffic. Figure 3 shows one ring interface, with the three add- and two transit queues. The numbers in the circles indicate a crude priority on the output link. Regarding priority between add traffic and the STQ, as the STQ fills up, it will have increasingly higher priority (this is not a linear function, but bad on thresholds). Since class A frames have priority over all other
>宝宝大便有泡沫
