The major resources of concern in a bridge are the buffering required to store and forward frames , the table space for the forwarding databas e , and the CPU cycles to execute the algorithm. Note that CPU cycles are also required to perform network management. Typi cally, any bridge implementation must guarantee that network management commands are eventu ally executed . For example, suppose a bridge was heavily loaded because of a slow outbound
IAN. A network manager wanting to disconnect that bridge may be unable to do so if all received
frames are being drop
p
ed because of buffer congestion. Therefore , one important aspect of implementing a network management architec ture is that some amount of buffering must be preallocated to handle those messages . More over, scheduling must be accomplished so that the network management process in the bridge is guaranteed to make progress. This guarantee is a matter of correctness and therefore should be stated in any effort to make the architecture a standard.
Buffers are also required to hold frames while they are waiting to be either processed or for
warded. As depicted in Figure 3, bridge can be
modeled as a queuing system in which the ser vice centers represent the forwarding process and the outbound LANs. Congestion can occur at three places:
1 . Upon reception, owing to the lack of receive buffers
2 . After reception, owing to queuing for the forwarding process
3. After the forwarding process, because of con
gestion on the outbound LAN
Proper bridge design can solve the first two sources of congestion. The third problem, how ever, is a general one for bridges, routers, and
any store-and-forward device. 20 There are several
ways that the bridge designer can address this problem. We first make a general observation about the required service rate of the service centers in a queuing network. Steady-state con gestion at the forwarding process can be avoided completely if the network can always make for warding decisions faster than the summation of the interarrival times of the smallest frames across all the inbound LANs. The forwarding database must be consulted for each frame on which a forwarding decision is made. There are many ways to do that very efficiently.
The table discussed earlier is really only a cache of station address-to-MAC entity associa tions; a search of that table is required to locate an entry. If the table is ordered, then a binary search can locate the entry in question. There are other alternative search methods, such as seg mented hashing. The implementation of this pro-
BRIDGE ,---�---� 64 I I BUFFERS I I I I
T
. DISCARD I I I I I I L---jFigure 3 The Two Port Bridge Resource Model
Digital TecbnicalJournal No. 3 September 1986
cess is one of the key aspects of the bridge tech nology. This facet is covered later in the paper in the discussion of the technology used in the IAN Bridge 100 product.
A final point with respect to caching is in order. Further enhancements in performance can be obtained by recognizing something about the nature of the traffic on IANs. Extensive measure ments on token rings, Ethernets, etc. have uncov ered several important facts. These are related to the nature of higher-layer protocol and applica tion operation. One is that, given that a frame from station S and station D has just been observed on the
LAN
, the probability that thenext frame observed is either from D to S or also from S to D is very high.21 Thus, if the bridge keeps the last few associations it has obtained from the database, it is very likely that the next frame will use one of those associations. Keep ing them further reduces table access rates. It amounts to a two-level cache.
With these observations we now focus on con gestion at the receive or transmit buffers. Con gestion at the receive buffers can- be avoided through proper machine organization. For exam ple, a bridge using separate controllers for each IAN, each controller having its own local buffer ing, will have to assure that sufficient buffering
APPLICATION LAYER
NETWORK LAYER
is available to maintain stability in the queue (particularly during transient bursts of frames) . Frames will have to be moved (out of the con troller buffers or shared m�mory) into another buffer to queue for the forw
�
I rding process. With respect to bridge delay, th�s time must also be included in the forwarding process. With respect to bridge throughput, the bottleneck server will determine the peak.Therefore, the only place any congestion will occur in these bridges is at the outbound IAN. This congestion will occur if that
LAN
is not fast enough for the volurrl
e of traffic it must carry. This problem is an iss�
e of IAN speed, not bridge speed. The philosophy is to design bridges so that they will not be bottlenecks. Most of these comments apply to any routing algorithm and hold true whether a table or a frame must be searched. AQ.d they hold true forall the MACs. ·
Effect of Bridges on Ethernet Links Bridges have several effects' on the performance of CSMAfCD IANs. One effect is due to the filter ing function that prevents traffic from entering a subnet that it need not traverse . Recall that bridges operate above the data link MAC layer as
shown in Figure 4. Preventing this traffic flow
APPLICATION LAYER NETWORK LAYER - - - · - - - -
j
_ _ _ _ _ _ _ - - - - DATA LINK LAYER PHYSICAL LAYER PHYSICAL MEDIUM BRIDGE BRIDGE FUNCTIONS DATA LINK DATA LINK LAYER LAYER PHYSICAL PHYSICAL LAYER LAYERPHYSICAL MEDIUM
Figure 4 Bridges and Data Links
DATA LINK LAYER PHYSICAL LAYER Digital Technicaljournal 65 No. 3 September 1986 New Products
The Extended Local Area Network Architecture and LAN Bridge 100
reduces the applied load on the LAN , thus improving performance for the local users.
Another effect is more subtl e . Consider a CSMAfCD system with an extent of D meters and N stations distributed uniformly over that extent. Without using bridges, all N stations have to share the resources of that one LAN extended over D meters. The delay and capacity are deter mined by the applied load, as described above. If added in the center of the system, the Ethernet will be partitioned into two Ethernets, each with N/2 stations . Thus the collision windows on each partition have been cut in half. The smaller collision windows cause less bandwidth to be wasted per collision. The net effect on the sys tem is not only to reduce the load applied to a given Ethernet (through filtering) , but also to improve the overall efficiency or capacity of that Ethernet since the extent it must cover is smaller. In effect, the Ethernet gets more efficient as the load applied to it is reduced. Given these factors, along with performance information characteriz ing the behavior of the Ethernets under load, the bridge designer can investigate the performance of bridged networks using these IANs.4,22
The remaining section of this paper discusses
the IANBridge 100, which is an implementation
of the Extended IAN Architecture.