cs398 Lecture Notes Spring 2000 Week 7, Tuesday For today, you should have re-read Section 3.3 and written answers to questions 21, 22 and 23. Limitations of bridges ---------------------- 1) they don't scale The MST algorithm takes linear time (a number of rounds that is proportional to the diameter of the graph) Broadcast becomes infeasible/undesireable. 2) can't handle heterogeneity performance heterogeneity not too bad protocol heterogeneity is harder a) make bridges smarter, so they can read/write multiple protocols b) encapsulate the forwarding/routing level information within network level frames (Chapter 4) 3) transparency is a two-edged sword multiple hops masquerading as a single hop is a nice abstraction, but it hides details that affect design decisions at higher levels 1) possibility of dropped frames in bridges 2) more variability in latency 3) possible reordering This last point explains one of the confusing things about Chapter 2, which is that many link level protocols contain features to deal with these problems; problems that don't seem relevant at the link level, unless the link level is implemented by a network of links and bridges! That's the end of the general comments in Chapter 3 about switches and bridges. There are two sections left: 3.3 ATM: a discussion of a specific switching design 3.4 Hardware: a discussion of hardware design of switches in general. ATM --- Asynchronous transfer mode Connection-oriented, using virtual circuits. Switches reserve buffer space and bandwidth, and so can make performance guarantees with levels of QoS Fixed length packets called cells. Advantages of fixed length 1) simpler hardware 2) facilitates parallelism 3) finer control over queueing behavior (priorities) 4) reduction in jitter (variability in delay) 5) less bandwidth lost due to store and forward (when queue is empty -- irrelevant on loaded network) Disadvantages of fixed length 1) overhead of cell headers 2) wasted bandwidth on padding small packets Cell size --------- Tradeoff between large and small cells Large -- lower overhead Small -- less padding, less store-forward delay Example: voice over ATM 64 Kbps = 8000 samples per second * 8 bits per sample Large cells mean noticeable delays 1000 bytes takes 125 ms to accumulate at sender before first cell leaves Human perception threshhold is about 100 ms = 0.1 s Cell format ----------- Drop priority -- why is this useful? Why is stream data like justice? The header is heavily protected by CRC (4:1!) and it uses error correction. Why? Segmentation ------------ ATM adaptation layer (AAL) intermediate protocol funny book, page 204 PDU = protocol data unit = AAL frame Two choice for data AAL3/4 -- puts 44 bytes of data and a 4-byte AAL header into each cell header identifies BOM, COM, EOM and SSM contains 10-bit CRC sequence number for losses or reorderings multiplexer for sharing a VC among data flows AAL5 -- uses all 48 bytes for data borrows one bit of the ATM header to identify EOM no CRC (per cell), no sequence number, no multiplexing BUT, each PDU gets a trailer that contains 32-bit CRC for the whole PDU. funny book, page 209 Virtual path ------------ 24 bit identifier, broken into 8-bit virtual path and 16-bit VC Distinction between public and private switches. Public use only VPI to switch all traffic between two sites. Private all 24 bits Faster, less state in network, less setup time. Physical layers --------------- Natural to run ATM over SONET because both used fixed size frames. Other physical media available. Framing problem: how to find beginning of cells. on ATM: 1) SONET header points into SONET payload 2) check to see whether the 5th byte is a CRC of the previous four ATM vs. Ethernet ---------------- At several points in time, ATM has been running on faster physical media than Ethernet, but 1) it is not clear whether this is accidental or intrinsic 2) the advantage has been short lived Naively, switched networks have an advantage over shared media because in a shared medium the total bandwidth is divided among the hosts, and in a switched network, each host gets full bandwidth to the switch. But this description is misleading, since this distinction is not intrinsic to the technology, it is a description of the typical configuration of each. An Ethernet with one host per subnet is equivalent to a switched network, except that it is more flexible, since you have the option of allowing hosts to share a link, performance requirements permitting. So the real issue is, how much does it cost to set up these configurations (which includes ease-of-use)? The only _intrinsic_ difference is that ATM does not support broadcast. That means the ARP we discussed doesn't work. Have to use ATMARP, or make the switch broadcast (without slowing it down). Performance issues ------------------ General-purpose hardware may not be appropriate for switches 1) bandwidth limit: half the speed of the I/O bus or half the memory bandwidth, whichever is less 2) throughput limit: throughput = number of packets per second * packet size for small packets, often limited by software speed (reading headers, looking up routing table, etc.) (this is the first definition we have seen. note that it is not the same as bandwidth -- a switch that is throughput limited will not saturate the bandwidth of the outgoing links)