Carrier Extension Issues Stephen Haddock Extreme Networks 5/21/96 Introduction: Scaling 802.3 MAC to 1Gb/s without modifying the slot time results in an extremely small network diameter. The carrier extension proposal allows a larger network diameter by increasing the slot time from 64 to 512 byte intervals. Increasing the slot time has several undesirable side effects. This presentation examines those side effects and proposes some mechanisms to mitigate them. Implications of 512 byte slot time: -- Larger transmit and receive FIFOs. -- End of packet may precede end of carrier. When is packet transmission successful? -- Reduced network efficiency: When transmitting short packets and When collisions occur. -- Amplification of capture effect Larger Tx/Rx FIFOs: -- Transmitter must be prepared to retransmit up to 512 bytes. -- Receiver must be prepare to flush up to 512 bytes. Whole Packets within Fragments: Station A transmits short packet with carrier extension. Station B (at maximum distance from station A) collides with A. Fragment received by station C may contain the entire packet but must still be discarded to avoid duplication of data when A retransmits. Network Efficiency: Efficiency = Useful Data / (Useful Data + Overhead) For calculations below, "useful data" includes the 802.3 data field, and "overhead" includes IPG, preamble, MAC header, and FCS. Carrier Extension on small packets increases overhead. Collisions reduce efficiency due to: Time spent in collision itself. Idle time when all stations in backoff. Network Efficiency: Small Packets Packet distribution Efficiency 64 byte 512 byte slot time slot time ---------------------------------------------------- All minimum size packets 55 % 9 % 50% min and 50% max 95 75 Workgroup Avg distriburion 94 72 Network Efficiency: Collisions Collision frequency Efficiency (Workgroup Avg packet 64 byte 512 byte size distribution) slot time slot time ---------------------------------------------------- No collisions 94 % 72 % Mohan's simulation of 86 56 15 station, 200 m network Capture Effect: Capture Effect is a characteristic of the Binary Exponential Backoff algorithm where, given one or more stations capable of generating extended streams of back- to-back packets, the station that first wins a collision backoff gains an advantage that increases with each successive collision. This results in a station being able to "capture" the media and transmit a large number of consecutive packets while other stations are effectively locked out. Results of Capture Effect: -- Transient unfairness in media access. -- Extremely variable, and sometimes extremely long, media access latencies. -- Long periods where network is idle because all stations in backoff. -- Packet loss due to excessive collisions. -- Packet loss due to queue overflows. -- Negative interaction with transport protocols. Capture Effect vs. Slot Time: Mohan Kalkunte's simulation data for number of consecutive packets transmitted vs slot time. [ Chart showing mean, std dev., and 95th% number of consecutive packets at slot times of 64, 128, 256, and 512 bytes. Number of consecutive packets transmitted is proportional to slot time, and the 95th% number increases from about 120 packets at a 64 byte slot time to just under 1000 packets at a 512 byte slot time. ] The amplification of the capture effect is NOT a direct result of scaling the MAC. It is a side effect of the carrier extension! Why worry about this now? Capture Effect becomes most significant on networks with a small number of stations that are each capable of offering a heavy traffic load. The trend toward mixed speed switching (10/100 and 100/1000) funnels traffic to the higher bandwidth networks creating the worst case configuration for Capture Effect. Even brief periods of congestion can cause very long and unpredictable queuing delays. The trend toward interactive applications on the network creates data that is the most sensitive to Capture Effect. Potential Solutions: -- Don't share it, switch it. Abandon a multi-node shared network topology altogether. -- Don't use carrier extension Accept a very small network diameter. -- Buffered repeater proposal Modify repeater architecture so carrier extension is not required. -- Optimized carrier extension proposal Enhance the operation to mitigate effects of longer slot time. Buffered Repeater: -- Concept: Move the retransmit buffer from the DTE to the repeater. -- Results: Entire collision domain is inside the repeater. Slot time becomes arbitrarily small. Buffered Repeater: Advantages -- DTE transmitter never needs to retransmit. -- DTE receiver never sees a fragment (potentially). -- No need to extend carrier on short packets. -- Negligible loss of efficiency due to collisions (potentially). -- Capture effect no worse than on 10/100 Mb/s Ethernet (and perhaps much better). Buffered Repeater: Disadvantages -- Limited to single repeater topology. -- Increased repeater cost/complexity. -- Observable changes in MAC behavior e.g. MAC no longer has visibility to collisions so cannot maintain collision counters. -- Limits max. packet size. Larger packet size is not an objective of 802.3z, however there is nothing else in MAC/repeater architecture that enforces a maximum size. Enhanced Carrier Extension: Step 1: Allow multiple packets within a carrier event. Minimal incremental burden on MAC Tx/Rx. Virtually eliminates efficiency loss on small packets. Step 2: Limit capture effect. Several proposals in the industry. Very important to make it at least no worse and hopefully better than current 10/100 Mb/s Ethernet. Multiple Packets within Carrier: Tx Rule 1: Transmitter may begin a new packet if current carrier duration is less than the slot time (512 bytes). -- Must keep at least min. IPG between packets. -- Results in up to 7 packets within carrier event. -- Max. carrier event approximately 2 KBytes (slot time plus max. packet) -- PCS encoding must allow a (properly aligned) S after R without intervening idle. Tx Rule 2: If a collision occurs, the transmitter must retransmit all packets included in the carrier event. Rx Rule: Receiver must discard all packets received within a single carrier event if total carrier duration is less than the slot time. Options to Limit Capture Effect: -- Scale down maximum backoff time. -- BLAM -- Capture Avoidance Binary Exponential Backoff (DEC) -- Other capture avoidance algorithms (?) Scaled Down Backoff Time: Backoff time is a number of slot time intervals r, where r is a uniformly distributed random number constrained by: 0 <= r < 2**k ; k = min (n, 10) Compensate for a factor of 8 increase in slot time by modifying this to: k = min (n,7) This compensation has no effect for the first 7 collision attempts, but thereafter limits the maximum backoff to the same number of bit times as in 10/100 BaseT. Should effectively limit the 3-sigma value of the capture effect to roughly the same as 10/100 BaseT. May compromise the ability of the BEB algorithm to resolve contention between more than 2**7 stations. Other side effects? Very simple solution, though at best it matches 10/100 BaseT performance with respect to capture effect. BLAM is more effective. Conclusions: Scaling the Ethernet MAC parameters to 1 Gb/s results in an extremely small network diameter. Adding a carrier extension to short packets and increasing the slot time to 512 bytes has been proposed as a means of maintaining a 200 meter network diameter. This has a negative impact on the network efficiency, and on the Capture Effect. If we pursue the extended carrier solution, then we should take steps to correct these two issues. 1) Carrier extension on short packets causes a significant loss in network efficiency: from 94% to 72% for the workgroup average packet size distribution. This efficiency can be regained by allowing multiple packets per carrier event. 2) Increasing the slot time by a factor of 8 amplifies the Capture Effect by a factor of 8. The increase in slot time can be compensated by reducing the maximum backoff time from 2**10 to 2**7, thus limiting the Capture Effect to the same relative behavior as in 10/100 Mb/s Ethernet.