Tables 9a and 9b provide an overview of the major architectural features of the switches included in the Switched 10 Mbps-100 Mbps Evaluation. Having a basic understanding of the switch architecture is often helpful as background information while reviewing the vendor-specific test results or the comparative test results. A brief explanation of each column entry follows the tables.
Shared Active Switch Intern Forwardi Bandwidt Dedicated Dedicated Buffer Congestio Fabric al ng Mode h of Input Output Pool n Data Switch Buffer/Po Buffer/Po Across Control Format Fabric rt rt Ports Mechanism 3Com Shared Frames Store-&- 360 Mbps None None 4M No LANplex Memory Forward 2500 3Com Shared Frames Store-&- 540 Mbps None None 4M No LANplex Memory Forward 6004 3Com Shared Frames Store-&- 100 Mbps 64 KB 64 KB None No LinkSwitc Medium Forward h 1200 3Com Shared Frames Store-&- 180 Mbps None None 1-4M No LinkSwitc Memory Forward h 2200TABLE 9A. ARCHITECTURAL FEATURES OF ETHERNET-TO-FDDI SWITCHES
Shared Active Switch Intern Forwardin Bandwidt Dedicated Dedicated Buffer Congestio Fabric al g Mode h of Input Output Pool n Data Switch Buffer/Po Buffer/Po Across Control Format Fabric rt rt Ports Mechanism 3Com Shared Frames Selectabl 800 Mbps 8 KB (10 24 KB None Yes, LinkSwitc Medium e;Cut-thr Mbps) (10 Back h 1000 ough, 64 KB Mbps) Pressure Fragment- (100 64 KB free, Mbps) (100 Store-&-F Mbps) orward Bay Shared Frames Store-&-F 2 Gbps 128 KB 128 KB None Yes, LattisSwi Medium orward Flow tch 28115 Control between LS 28115s Cisco Shared Frames Selectabl 1 Gbps None None 3 MB No Catalyst Memory e; 2100 Cut-throu gh, Fragment- free, Store-&-F orward Cisco Shared Frames Store-&-F 1.2 Gbps 32 KB 160 KB None No Catalyst Medium orward 5000 Cisco Shared Frames Adaptive; 520 Mbps 96 KB 96 KB None No Catalyst Medium Cut-throu 3000 gh, Fragment- free, Store-&-F orward CrossComm Shared Frames Store-&-F 2.4 128 KB 384 KB None No CS3002 Medium orward Gbps (10 (10 (640 Mbps) Mbps) Mbps 256 KB 768 KB per (100 (100 module) Mbps) Mbps) IBM Shared Frames Adaptive 512 Mbps 96 KB 96 KB None No 8271 Medium Cut-throu Nways gh, Ethernet Fragment- LAN free, Switch Store-&-F orward Madge Shared Cells Store-&-F 1.28 32 KB 32 KB None Yes, LANswitch Medium orward Gbps Back Pressure NBase Shared Frames Store-&-F 200 Mbps 6 KB 24 KB None Yes, NH208 Medium orward Back MegaSwitc Pressure hTABLE 9B. ARCHITECTURAL FEATURES OF ETHERNET-TO-FAST ETHERNET SWITCHES
Most switch architectures can be classified as either cross point matrix, shared memory or shared medium. Cross point matrix switches employ an array of switching elements to provide parallel switched paths between distinct pairs of input and output ports. This design approach has yielded fairly attractive price per port in Ethernet switches with relatively few ports. Shared memory architectures are also very common for low cost, small-scale switches and have the advantage of easily accommodating mixed LAN types and speeds within a single switch. Shared media switches use a high-speed backplane to interconnect switching elements, which may consist of an individual bridge per port or a multiport switch module. The latter may use shared media or shared memory as an internal architecture. Shared media architectures are frequently used to build modular switches that can scale to high port densities.
Internally the switch can use either the native Ethernet or Ethernet and FDDI frame formats or it can convert the frames into fixed length cells. While fixed length cells can simplify design of buffering and facilitate the addition of higher speed ports including ATM, the frame-to-cell and cell-to-frame conversion for LAN-to-LAN switching does incur additional processing overhead versus simpler frame switches.
Switches can forward packets in either store-and-forward, cut-through or fragment-free modes. Store-and-forward allows for full error checking, packet filtering and LAN speed conversions at the cost of higher transit delay, especially for large packets. Cut-through minimizes transit delay by foregoing the possibility of error checking and packet filtering. Fragment-free mode is cut-through switching in which runt packets (collision by-products of less than the minimum legal packet size) are discarded. Some switches allow the network manager to select the mode of Ethernet-to-Ethernet switching among the three possibilities. Other switches use cut-through as the default Ethernet-to-Ethernet switching mode but automatically adapt to fragment-free or store-and-forward mode when the frequency of damaged packets exceeds a preset threshold. Switches in cut-through mode must store the entire packet in switching from 10 Mbps to 100 Mbps.
For switches in all three architectural classifications the aggregate bandwidth of the switch fabric provides an overall measure of the capacity of the switch to support multiple simultaneous streams of Ethernet traffic or a mixture of Ethernet and Fast Ethernet or FDDI traffic. This specification is closely related to the ability of the switch to achieve high aggregate throughput.
Buffer space can be allocated in a number of ways. Three of the possibilities are: 1) to dedicate a fixed amount to each input port, 2) to dedicate a fixed amount to each output port or 3) to allocate space as needed from a common pool shared by input or output ports within a switch module. While these approaches have differing degrees of complexity, the best measure of their effectiveness is probably found in the results of the congestion tests performed in the evaluation.
Congestion control involves the ability of the switch to deal with oversubscribed output ports--a number of input ports contending simultaneously for a single output port or a single high-speed port forwarding traffic to a single low-speed output port. Congestion at the output ports can also occur when the output port is trying to forward traffic over a highly congested shared segment. Passive congestion tolerance is based on buffers holding packets in queues until the output port becomes free. Active congestion control techniques apply back pressure to the traffic sources by forcing actual collisions, using the Ethernet jam signals to spoof the occurrence of collisions, or raising the carrier signal to delay transmission by the end systems.
Tables 10a and 10b focus on general features of the switch including several that determine whether a particular switch may be better suited for workgroup or collapsed backbone applications.
Max Fault Switch Max. Full # of Full Number Toleran Packet Bridge Price Price Config # Duple High-S Duple of MAC ce: Filter / / ura-ti of x peed x Address Redunda ing Routin Ether on Ether Ether Ports High- es nt g net net net Speed Power Protoc Port Ports Optio Ports & Hot ols n Swap Modules 3Com Expand 16 No 2 FDDI No 8,192/ Redunda SA/DA, 802.1d $14,39 $900 LANple -able System nt M'cast ; 5, 2 x 2500 Power; , Transl FDDI Hot B'cast a-tion Ports Swap , al Protoc Bridgi ol ng; IP /RIP; IP frag 3Com Chassi 48 No 6 FDDI No 8,192/ Hot SA/DA, 802.1d $31,30 $978 LANple s Module Swap M'cast ; 0, 2 x 6004 Only , Transl FDDI B'cast a-tion Ports , al Protoc Bridgi ol ng; IP /RIP; IP frag 3Com Module 9 No 1 FDDI No Etherne As DA, Transl $6,995 $1,16 LinkSw for Chass t: Part M'cast a-tion , 1 6 itch Chassi is 1,024/ of , al FDDI 1200 s Hub or Switch; Stack B'cast Bridgi Port or 6 FDDI: or ng Stand- Stand No Chassis alone -alon Limit Hub e 3Com Fixed 16 No 1 FDDI No 8,192/A As SA/DA, 802.1d $9,995 $625 LinkSw Config ll Part M'cast ; , 1 itch Stacka Ports of , Transl FDDI 2200 ble Stack B'cast a-tion Port , al Protoc Bridgi ol ng; IP fragSA/DA = Source Address/Destination Address
* Network ports, either Ethernet or Fast Ethernet have unlimited MAC addresses per port. TABLE 10A. GENERAL FEATURES OF ETHERNET-TO FDDI SWITCHES
Fault Switch Max. Full Max Full Number Toleran Packet Bridge/ Price Price Config # of Duple # of Duple of ce: Filter Routing / ura-ti Ether x High- x MAC Redunda ing Protoco Ether on net Ether Speed High- Addres nt ls net Ports net Ports Speed ses Power Port Optio Ports & Hot n Swap Modules 3Com Fixed 24 No 2 Futur Ethern As DA, $4,975 $207 LinkSwi Config 100BT e et: Part M'cast - 1 tch , X 500; of , 100BTX 1000 Stacka 100BT: Stack B'cast Port ble Unlimi ted* Bay Expand 16 No 18 FE Yes 1,024/ Redunda SA,DA, LattisS $18,95 $1,14 LattisS -able, port nt M'cast pan 0 0** witch Stacka Power , proprie 28115 ble B'cast tary Cisco Fixed 25 No 2 Yes 1,024/ Hot SA, 802.1d $6,295 $252 Catalys Config 100BT Port Swap DA - 2 t 2100 . X M'cast 100BTX , Ports B'cast Cisco Chassi 96 Yes 50 FE Yes, 16,000 Yes 802.1d $42,00 $875* Catalys s FE / Port 0 - * t 5000 14 100BTX Ports Cisco Expand 24 Yes 2 Yes 1,700/ No SA/DA, 802.1d $9,690 $605 Catalys -able, 100BT Port M'cast - 1 t 3000 Stacka X , 100BTX ble B'cast Port , CrossCo Chassi 64 Yes 4 Yes 4/Ethe Yes SA/DA, 802.1d $7,895 $263 mm CS s, 100BT rnet M'cast 3002 Stacka X Port , ble B'cast , IBM Expand 12 Yes 1 Yes 1,700/ No SA/DA, 802.1d $5,300 $662 8271 -able 100BT Port M'cast - 1 Nways X , 100BTX Etherne B'cast Port t LAN , Switch Madge Chassi 128 Yes 16 No Ethern Yes SA/DA, $55,150 $920 LANswit s 100BT et: VLANs - 1 ch or 8 64,000 100BTX FDDI /8 Port Ports; 100BT: 8/Modu le NBase Expand 6 Yes 2 Yes Standa Power Source 802.1d $5,295 $883 NH208 -able 100BT rd: Only Port, - 1 MegaSwi X or 1,024/ DA, 100BTX tch 100BF Switch M'cast Port X ; , Option B'cast al: , 2,048/ VLANs Switch* = Network ports, either Ethernet or Fast Ethernet have unlimited MAC addresses per port.
** = the Catalyst 5000 tested configuration includes 14 Fast Ethernet ports, inflating the price per Ethernet port considerably. With two Fast Ethernet ports, the Catalyst 5000 price per Ethernet port is $666.
The Bay 28115 Ethernet ports can be configured for either 10 or 100 Mbps speed, making the price per Ethernet port a less meaningful metric for this switch. TABLE 10B. GENERAL FEATURES OF ETHERNET-TO-FAST ETHERNET SWITCHES
The following is a brief discussion of each of these general features.
Ethernet switches are available in expandable, stackable, modular or fixed configuration formats. Most of the expandable and stackable switches in the chart have the ability to accommodate a very limited number (one to three) expansion modules for additional Ethernet or individual high-speed ports. Some of the smaller switches can also be used as modules in enterprise hubs. Such a switch may function as a "collapsed backbone" switch connecting shared media segments within the hub or as a means of converting the hub into a modular Ethernet switch for external segments. By the same token, a modular switching hub may accommodate shared media modules and/or high-speed LAN modules to assume the role of a high performance enterprise hub.
Expandable, stackable and fixed configuration switches are generally most appropriate for small workgroup and small collapsed backbone configurations, while modular and hub module switches can scale up for larger workgroup and backbone applications.
The maximum number of Ethernet ports that can be configured is an important consideration for switches that will be used in larger workgroup applications.
Full duplex Ethernet is one option for increasing the bandwidth per port of Ethernet switches. Full duplex Ethernet eliminates collisions in dedicated Ethernet connections and can essentially double the bandwidth for links that support symmetrical traffic flow. Therefore, full duplex offers some reasonable benefits for switch-to-server connections and for switch-to-switch connections. The lack of availability of full duplex network interface cards (NICs) has limited the applicability of the technology for end system connectivity. Furthermore, since client/server interactions do not often involve symmetrical traffic flows, full duplex does not have as much impact on the effective bandwidth in end system-to-switch connections.
Maximum Number of High-Speed Ports
FDDI, 100Base-T, 100VG-AnyLAN and ATM all constitute options as "fat pipes" for connecting switches to servers and/or to high-speed backbones. All of these LAN technologies feature at least 10 times the bandwidth of Ethernet and are also capable of full duplex operation in all cases except for 100VG-AnyLAN, which is a half duplex, shared media LAN technology. With today's mix of installed end systems and applications there is only limited need for 100 Mbps switched desktop connections. Future generations of Ethernet switches may offer 100 Mbps ports for desktop connectivity and 622 Mbps ATM or Gigabit Ethernet connections to servers and backbones.
Full duplex operation is also an option for increasing the effective bandwidth of high-speed switch ports. Full duplex Fast Ethernet eliminates collisions in dedicated Ethernet connections and can essentially double the bandwidth for links that support symmetrical traffic flow. Full duplex FDDI also doubles two-way bandwidth for point-to-point connections. Therefore, full duplex high-speed ports can reduce bottlenecks for switch-to-server connections and for switch-to-switch connections.
The amount of memory devoted to address tables can limit the number of MAC addresses that can be stored per port. Switches with a maximum of a single Ethernet MAC per port have been cost-optimized for "private Ethernet" desktop connectivity. Switches with a limited number of MACs per port cannot be used in any sort of collapsed backbone or LAN segmentation configuration involving the interconnection of segments that are shared by a large number of stations. Most of the switches in the evaluation have the capacity to store over 100 MAC addresses per port, and a number of switches can even store 1,024 addresses per port. This provides somewhat more flexibility in how the switches may be configured generally at the cost of higher price per port.
The number of MAC addresses supported on the switch's high-speed ports is another important specification. Some Ethernet/Fast Ethernet switches intended for workgroup applications perform only simple non-802.1d bridging and forward all traffic destined for unknown addresses over the high-speed port intended for backbone network connectivity. The assumption here is that the switch will quickly learn the addresses of all the stations on its local ports dedicated to individual desktops, servers or small shared segments. On the other hand, switches intended for segmentation and backbone applications generally maintain address tables for all ports of the switch, often with larger table capacity for high-speed backbone ports.
Fault tolerance is usually provided in the form of redundant power supplies and/or hot swappable modules (in the case of modular devices). Switches that have been designed as modules for chassis-based enterprise hubs can themselves be hot swappable units within the hub chassis and can draw their power from the hub's redundant power supply. In other words, the switch module essentially inherits the fault tolerant features of the host hub. In a similar fashion, a stackable switch can function a hot swappable module of a stack drawing backup power from a power supply shared with the rest of the stack. Fault tolerance becomes an important issue where a large number of users would lose access to the network in the event of switch failure. Therefore, most of the switches with high port count or intended for backbone applications can be expected to have native or "inherited" fault tolerant features.
Most store-and-forward switches have the capability to do some form of packet filtering beyond the simple discarding of damaged (runt) or misaligned packets. Packet filtering based on source or destination address or protocol type can be used as a security measure or to exclude unwanted traffic from being forwarded over the backbone or to secure LAN segments. Some measure of protection from broadcast storms can be achieved with a switch that is capable of filtering out broadcast and multicast packets that exceed a threshold level that the network manager can control. Packet filtering is obviously more important for backbone applications, but many workgroup switches include this capability as a means of preventing unwanted traffic from crossing the boundary (in either direction) between the workgroup and the backbone.
Most Ethernet switches include support for the IEEE 802.1d Spanning Tree protocol. This allows