index.md (3161B)
1 +++ 2 title = "Transport: elements of transport protocols" 3 +++ 4 5 # Transport: elements of transport protocols 6 **Elements of transport protocols** 7 Addressing 8 9 - NSAP (network service access point) — IP address 10 11 - TSAP (transport service access point) — port 12 13 - some services are always on the same ports (e.g. mail server on TCP 25, ssh on TCP 22) 14 15 - portmapper process 16 17 - finds TSAP port for a specified service name, using its database 18 19 - a new service registers itself with the portmapper 20 21 - kind of like directory assistance phone operator 22 23 - initial connection protocol 24 25 - each machine that offers services has a process server (inetd on UNIX) 26 27 - this acts a proxy for less used servers 28 29 - listens to set of ports at same time, waiting for a conn request 30 31 Connection establishment 32 33 - using sequence numbers for packets 34 35 - label packets with sequence numbers that will not be reused within T seconds 36 37 - packet hop limit used to remove old packets 38 39 - time-of-day clock used to choose sequence number 40 41 - three-way handshake 42 43 - one peer checks with other if connection request is current 44 45 - e.g. host 1 chooses sequence number x and sends it in CONNECTION REQUEST to host 2, host 2 replies with ACK containing its own sequence number y, host 1 acknowledges host 2’s sequence number in first data segment it sends 46 47 Connection release 48 49 - asymmetric release — one host gtfo, other host is left hanging. can lead to data loss. 50 51 - symmetric release — participants agree to end connection 52 53 - two-army problem — one white army vs two blue armies. one white beats one blue, but two blue beat white if attack at same time. can they coordinate attack using messengers? No. 54 55  56 57 - one way to semi-solve this is by timers and only allowing a certain number of timeouts, but y’know…it’s not great 58 59  60 61 Flow control 62 63 - slow down sender if receiver can’t handle the data rate 64 65 - received packets are buffered at the receiver 66 67 - buffer management is done separately 68 69 - available buffer space is receiver window size 70 71 Crash recovery 72 73 - recovery from layer N crash can only be done by layer N+1 (higher abstraction) 74 75 - but only if layer N+1 has info that lets it reconstruct previous state 76 77 - e.g. if there’s a problem on the transport layer, the transport layer’s lazy ass leaves it to the application layer to fix its problems 78 79 Congestion control 80 81 - needed to slow down sender if network can’t handle the data rate 82 83 - bandwidth allocation 84 85 - max-min fairness 86 87 - maximises minimum bandwidth, then uses excess where possible 88 89 - bandwidth given to one flow can’t be increased without decreasing bandwidth to another flow 90 91 - amount of bandwidth needed for a connection varies, so need to converge to ideal point 92 93 - sharing bandwidth between users 94 95 - dynamic adjustments with AIMD — convergence: 96 97  98 99 - problem is, TCP data loss means congestion, and wireless loses data commonly. so wireless must mask data loss