Transmission Control Protocol, ATM Network Layer

TCP algorithms:
- Sender/Receiver Silly Window Syndrome
- Jacobson timeout
- Karn selection of RTT estimates
summary of TCP congestion and flow control
ATM Network Layer
- ATM cell format
- ATM connection establishment
- ATM forwarding
- ATM service categories
- ATM Quality of Service (QoS)
- ATM congestion control
- ATM LANs

Silly Window Syndrome

I am sending a 5000-byte buffer with an MSS of 540
the last segment only has 100 bytes of data
then I start sending the next buffer
when receiver acks that last segment, my window will only increase by 100 bytes
if I am being held up by the flow control window, I can only send a 100-byte segment!
avoidance:
- sender side: hold off sending until we have 1 MSS or 1/2 the maximum receiver window
- receiver side: hold off sending ack until we have received 1 MSS or a timer expires

Jacobson/Karels algorithm
keep a running average of the round-trip time:
1. keep track of time t₀ at which we sent packet i
2. if the ack for that packet returns at time t_r, the new round-trip-time estimate is RTT_i = t_r - t₀
3. the new average is RTT_ave = alpha RTT_ave + beta RTT_i (typically, alpha = 0.875 and beta = 0.125)
also keep a running average of the RTT deviation: d_ave = alpha' d_ave + beta' |RTT_i - RTT_ave| ( alpha' = 0.75 and beta' = 0.25)
timeout = RTT_ave + 4 * d_ave

if I send a segment at time t₀, retransmit it at time t₁, and get an ack for it at time t₂, is my RTT estimate t₂ - t₀ or t₂ - t₁?
we can't tell
Karn: for timeout estimation, only use packets that were not retransmitted

Sender can only send when it has space in the window
window can be reduced by receiver (window carried in every ack segment)
Sender reduces its window for congestion control if packets are dropped
"slow" start: start at one segment, add one segment to the congestion window for each segment acknowledged, up to a threshold
linear increase: add one segment to the congestion window for every window acknowledged
multiplicative decrease: if a packet is lost, cut the threshold to 1/2 the current window and go back to slow start

routing happens at connection setup time:
- identify route
- reserve resources (if any)
- set up forwarding tables
forwarding table:
incoming incoming outgoing outgoing
port VPI/VCI port VPI/VCI
1 99 4 14
1 93/77 4 15/82
4 14 1 99
...
forwarding can be done in hardware: fixed-size cells, simple table lookup

CBR: constant bit rate (voice, uncompressed video)
VBR: variable bit rate (compressed video, other real-time and non-real-time applications)
ABR: available bit rate (PCR/MCR specification, "slow down" feedback from network)
UBR: unspecified bit rate (loss is acceptable, higher layers will deal with it)

leaky bucket outflow is SCR, capacity is CDVT
unless there is congestion, forward non-conforming cells, but with CLP = 1
Resource Management (RM) cell can be sent by sender, travels along route, records maximum bandwidth to use

different solutions
- every host has a connection to a LAN server, which broadcasts to all other hosts
- every host has a connection to an ARP server (ATMARP), which provides IP to ATM address translations. Each host can then establish an ATM connection to every other host in the LAN with which it communicates
VLAN (Virtual LAN): some hosts on the same ATM network might be using different ATMARP server, they "look" like they are on different LANs (simple-minded security)