Transport Layer. Application / Transport Interface. Transport Layer Services. Transport Layer Connections

Transcription

1 Application / Transport Interface Application requests service from transport layer Transport Layer Application Layer Prepare Transport service requirements Data for transport Local endpoint node address + service port Remote endpoint node address + service port Receives Application layer PDU data for transport Interface Control Information (ICI) Service requirements Local + remote endpoints Computer Networks Hadassah College Fall Computer Networks Hadassah College Fall Services Connection management Maintain separate sessions between various user applications Label sessions with client / server port numbers Provide reliability services according to connection type Multiplexing Divide outgoing data stream into segments Combine segments from multiple sessions into single output Demultiplexing Separate incoming segments by sessions Combine incoming segments for each session into user data stream Connections Reliable transport (TCP) Connection-oriented TCP connection established before data transfer Error-free delivery Data delivered In original order No s, duplications, omissions Flow control Control sender rate to prevent buffer overflow in receiver Congestion control Control sender rate to prevent buffer overflow in network Unreliable transport (UDP) Connectionless Lower overhead faster but no guarantees Segments with s discarded with no warning to application Computer Networks Hadassah College Fall Computer Networks Hadassah College Fall 4

2 5 Source / Destination Ports Client Opens socket to send requests Client / OS binds port number to socket 4 client port 65,55 identifies client application Server Opens listen socket mapped to accept sockets for requests Binds well-known port to service socket well-known port identifies service application Multiplexing / Demultiplexing Applications send / receive data on sockets Multiple sockets multiple conversations Transport layer segment Transport header + application data (PDU) TCP / UDP headers carry source + destination ports Multiplexing / demultiplexing Segments transmitted on same infrastructure Sorted by destination port at destination Client Application Bind socket to port 5 Connect to port 8 Transport Response src: 8 dest: 5 Request src: 5 dest: 8 Server Application Bind socket to service port 8 Accept from 5 Transport Client Applications Transport Server Application Transport Computer Networks Hadassah College Fall Computer Networks Hadassah College Fall 6 User Datagram Protocol (UDP) Internet unreliable transport protocol Defined in RFC 768 Used when low delay / jitter more important than control Streaming multimedia, multiplayer games,... UDP segment header UDP header source port length bits application data destination port checksum Length Number of bytes in datagram < 6 = 65,56 Maximum length = 6 lengths of all headers UDP / TCP Checksum Calculation at source Break UDP segment into sequence of 6-bit words Add IP pseudo-header (IP src/dest addresses, protocol, length) Pad with zeros if necessary Add all 6 bit words (not counting checksum field) Add carry-out to -order bit (rotate) Perform 's complement Calculation at destination Repeat calculation and compare Example rotate sum checksum Computer Networks Hadassah College Fall 7 Computer Networks Hadassah College Fall 8

3 9 Using UDP Client Application Open socket Send data on socket to endpoint (node address + service port) Client UDP Agent Accept data Add header with checksum Send to server Server Application Open socket Bind service port Listen on socket Receive data from listen socket Server UDP Agent Perform checksum Error Discard segment No Pass data to socket by port TCP Reliable Data Transport TCP provides reliable transport over unreliable layers Network and infrastructure layers can introduce s Bit s ( ), lost packets, duplicate packets, out-of-order packets TCP detects and corrects s TCP service abstractions rdt_send() Local Application data reliable transport udt_send() Local Reliable transport unreliable network rdt_rcv() Remote Unreliable network reliable transport deliver_data() Remote Reliable transport application layer rdt_send udt_send Application TCP Network Reliable Transfer Unreliable Transfer Application TCP Network deliver_data rdt_rcv If required add reliability features at client / server application level Layer Physical Transfer Layer Computer Networks Hadassah College Fall Computer Networks Hadassah College Fall Basic Bottom-up approach Version. for -free channel No checking necessary Version. Handles bit s Version. Bug fix of version. Version. Simplification of version. Version. Handles missing packets Version 4. Improved performance Version. on -free channel sndpkt sndpkt Error-free channel Data delivered reliably Computer Networks Hadassah College Fall Computer Networks Hadassah College Fall

4 Handling Bit Errors Error detection adds checksum to packet header Re-computes checksum Compares with checksum in header Corrupt packets (packets with s) discarded Error control provides sender with feedback about received data Version. on channel with bit s sndpkt NAK Acknowledgement Negative acknowledgement (reject) Data received without bit s Data received with bit s sndpkt sndpkt NAK Automatic Repeat Request (ARQ) re-transmits on NAK Computer Networks Hadassah College Fall Computer Networks Hadassah College Fall 4 Bug in Version. Corrupt / NAK packet Option interprets corrupt / NAK = misses data packet Option interprets corrupt / NAK = NAK retransmits packet may receive duplicate packet Version. with SEQ sndpkt Bug fix Label packet with sequence number (SEQ) SEQ = or sufficient for stop and wait follows option gets duplicate packet Re-transmits Discards duplicate packet sndpkt sndpkt discards duplicate packet Computer Networks Hadassah College Fall 5 Computer Networks Hadassah College Fall 6

5 7 Complication in Version. Version. SEQ + + NAK or NAK = response to last packet Corrupt or NAK = NAK NAK re-transmit packet or NAK for each packet Corrupt packet re-transmit packet Duplicate SEQ re-transmit but discard packet Version. SEQ + without transmitted NAK + SEQ = response to packet SEQ Corrupt = to previous SEQ = implied NAK Implied NAK re-transmit packet + SEQ for each packet Corrupt packet re-transmit for previous SEQ Duplicate SEQ re-transmit but discard packet Version. with SEQ + without NAK sndpkt sndpkt sndpkt Implied NAK: indicates in packet by resending for packet Computer Networks Hadassah College Fall Computer Networks Hadassah College Fall 8 Lost Packets Packet loss Lost data packet or Discarded by intermediate system or lower network layer Buffer overflow At receiver no or indication of lost packet Version. with SEQ + + timeout lost Handling lost packets sets timeout counter = corrupt = to previous SEQ = implied NAK Duplicates handled by SEQ timeout sndpkt sndpkt lost discards duplicate packet Computer Networks Hadassah College Fall 9 Computer Networks Hadassah College Fall

6 Performance Problem with Stop and Wait Wait too long Round Trip Time (RTT) sndpkt sndpkt 6 8 bits/b TT = transmission delay = 5 ms Mbps Typical RTT > ms T T 5 ms T T utilization = = =.4 TT + network latency TT + RTT 5 ms + ms Improving performance transaction Send packet Receive Stop-and-wait protocol Send packet wait receive Finish transaction n begin transaction n + Transmitter idle while waiting for Pipelined protocol Begin new transaction before previous finishes Reduce transmitter idle time Source window N = window size Transmit N packets before stop-and-wait N packets in process ("in flight" or "in pipeline") at any time Computer Networks Hadassah College Fall Computer Networks Hadassah College Fall Pipelining with window size = Pipelined protocol handling method Round Trip Time (RTT) sndpkt sndpkt sndpkt sndpkt 4 Selective repeat Buffers packet until counter for each packet sender re-transmits "uned" packet s received packets Received in order passed to application Received out-of-order held in buffer Typical RTT > ms N TT 5 ms utilization = =.4 T + RTT 5 ms + ms T sndpkt 5 Window size = N "uned" packets < N Send N packets without stop sending N < receiver buffer size Computer Networks Hadassah College Fall Computer Networks Hadassah College Fall 4

7 5 Selective repeat Bug in selective repeat Sequence number (SEQ) k bit SEQ SEQ k (modulo k ) Sequence number (SEQ) k bits SEQ k (modulo k ) Window size N k window k = SEQ N = Window size N k window k = SEQ N = waiting waiting Error waiting lost lost lost cannot distinguish cases:. Re transmission of duplicates. New packets Solution long SEQ ( bits or more) Computer Networks Hadassah College Fall Computer Networks Hadassah College Fall 6 Pipelined protocol handling method Go Back N (GBN) Buffers packet until counter for oldest "uned" packet sender re-transmits all "uned" packets from buffer sends CUMULATIVE for last packet Implies for all previous packets No buffer for out-of-order packets Error or missing packet no packet or any subsequent packet Window size = N "uned" packets < N Send N packets without stop sending N < transmit buffer size Version 4. Go Back N Sequence number (SEQ) k bits SEQ k (modulo k ) Window size N k window waiting waiting k = SEQ N = Error Computer Networks Hadassah College Fall 7 Computer Networks Hadassah College Fall 8

8 9 TCP as Reliable Transport Protocol Connection-oriented Set up connection before data transfer Maximum Segment Size (MSS) 6 (including IP header) Error detection Checksum as in UDP Error control ARQ with + SEQ + timeout No corrupt, missing, duplicate, or out-of-order data Piggybacking send within data segment Pipelining Variable window size at sender and receiver GBN cumulative with optional selective repeat Flow /congestion control Dynamic window size control of sender utilization TCP Header HLEN not used HLEN (data offset) Reserved Flags Window size Urgent pointer Options source port checksum acknowledgement number () 4 bits bits 9 bits 6 bits 6 bits sequence number (SEQ) flags bits bits Options destination port window size urgent pointer Length of TCP header in -bit words Not used Control bits Number of bytes receiver can receive now Offset from SEQ points to last urgent data byte Options fields + padding for multiple of bits Computer Networks Hadassah College Fall Computer Networks Hadassah College Fall TCP Header Flags TCP Connection Set up Three-way handshake NS CWR ECE URG PSH RST SYN FIN ECN-nonce concealment protection Congestion Window Reduced (CWR) flag indicates receiving segment with ECE flag set ECN-Echo If SYN = peer is ECN capable If SYN = packet with Congestion Experienced flag in IP header received during normal transmission Urgent pointer field valid Acknowledgment field valid Push buffered data to receiving application Reset connection Synchronize sequence numbers No more data from sender Client SYN segment SYN flag = SEQ = random number x No data Server SYN- segment SYN flag = flag = SEQ = random number y = x + No data Client segment SYN flag = flag = SEQ = random number x + = y + May contain data Client Connection request (synchronize) SYN flag = flag = SEQ = x = SYN flag = flag = SEQ = y = x + SYN flag = flag = SEQ = x + = y + data Server Accept Computer Networks Hadassah College Fall Computer Networks Hadassah College Fall

9 TCP SEQ + Random SEQ used in -way handshake Initial sequence number (ISN) Prevents hosts from accepting counterfeit segments At end of handshake SEQ = ISN + Byte sequencing SEQ indicates cumulative data from host during session SEQ = previous SEQ + number of data bytes in previous segment = ISN + + data bytes sent in all previous segments Relative sequence number SEQ ISN = data bytes + = next (expected) SEQ First data segment after handshake Client SEQ = x+ = y+ data = 5 bytes SEQ = y+ = x+5 SEQ = x+5 = y+ data = 4 bytes SEQ = y+ = x+9 Server Simplified TCP Start Event Receive application data Receive = y SEQ = ISN + SendBase = ISN + Response Send segment with sequence number = SEQ Timer not running start timer Pass segment to IP SEQ = SEQ + length(data) Send duplicate uned segment with smallest SEQ Start timer y > SendBase SendBase = y uned segments restart timer Computer Networks Hadassah College Fall Computer Networks Hadassah College Fall 4 Simplified TCP Event Response TCP SEQ + Error Segment arrives -free and in-order SEQ = next expected SEQ All previous data ed Delayed Wait < 5ms for next segment No next segment send SEQ = 9 8 data bytes Segment arrives -free and in-order SEQ = next expected SEQ One+ segment pending Send single cumulative SEQ = 9 8 data bytes Segment arrives corrupt or out-of-order SEQ > expected next Send duplicate for in-order SEQ = New segment after out-of-order segment SEQ = expected next Send immediate SEQ = data bytes Computer Networks Hadassah College Fall 5 Computer Networks Hadassah College Fall 6

10 7 TCP SEQ + Lost TCP SEQ + Missed SEQ = 9 8 data bytes SEQ = 9 8 data bytes SEQ = 9 8 data bytes SEQ = data bytes = = discards duplicate packet SEQ = data bytes SEQ = 9 8 data bytes = = = discards duplicate packet SEQ = data bytes Computer Networks Hadassah College Fall Computer Networks Hadassah College Fall 8 TCP SEQ + Cumulative all previous bytes SEQ = 9 8 data bytes SEQ = data bytes SEQ = data bytes = = Selective Acknowledgment Option Selective (S) Permits for segments with gaps Option negotiated between hosts Defined in RFC 8 Example Last = 5 Send 8 segments 5 data bytes / segment Case First 4 segments received and last 4 dropped returns normal = * 5 = 7 No S option field Data Case 5 First segment lost and 7 segments received 55 6 For each segment receiver returns segment with 65 7 = 5 75 S option field with start + end Option Field Start End Computer Networks Hadassah College Fall 9 Computer Networks Hadassah College Fall 4

11 4 Setting Value TCP Connection Close > RTT RTT = round trip time = minimum time to receive too short Too many missed s too long RTT Waste too much time before re-transmission Method Measure RTT Estimate timeout = RTT + margin On each new update timeout = ( - α)* timeout + α * measured_rtt Typical: α =.5 SEQ Symmetric Client or server may close connection FIN segment SYN flag = SEQ = cumulative SEQ number segment flag = = SEQ + FIN segment FIN flag = SEQ' = cumulative SEQ number segment flag = = SEQ' + Client FIN flag = SEQ flag = SEQ+ FIN flag = SEQ' flag = SEQ'+ Server Computer Networks Hadassah College Fall Computer Networks Hadassah College Fall 4 Flow Control and Congestion Control Flow control avoids overflow of receiver buffer Congestion control All sender avoid overflow of intermediate network buffers Buffer fill rate Bytes / second arriving from network Buffer empty rate Bytes / second leaving to network or application layer Buffer file time T overflow Example T overflow buffer size = buffer fill rate buffer empty rate Arriving bytes 6 6 = = = 6 seconds 8 KB/sec /sec /sec Empty Full Leaving bytes Flow Control Source window Initial source window = maximum number of "uned" bytes Determined by congestion + flow control Destination window Number of bytes receiver can accept Determined by available space in receiver buffer Buffer level = Previous level + arriving bytes bytes read by App Application reads too slowly decrease destination window Sliding window Arriving Windows field in TCP header bytes Number of bytes receiver will accept discards bytes above window size Empty Full Bytes read by App Computer Networks Hadassah College Fall 4 Computer Networks Hadassah College Fall 44

12 45 Flow Control Example Src Window KB 6 KB Persist uned KB KB KB 6 KB Computer Networks Hadassah College Fall + = 4 6 KB +4 = 6 KB 6+6 = KB + B KB KB window = KB window = 6 KB 6 KB window = KB B window = Buffer Level KB 6 KB KB App reads KB 6 KB 8 KB App reads Dest Window 8 KB 8 KB 6 KB Receive Window Issues Deadlock advertises window = Window update with window > is lost deadlock Persist timeout attempts small segment contains new window size Silly Window Problem Application reads received data slowly advertises small window Data bytes ~ header bytes More segments / file transfer larger total traffic (data + headers) Nagle Algorithm bug fix for Silly Window accumulates application data sends large segments Works badly with Telnet (requires small segments) side bug fix keeps window size until it can advertise large window Computer Networks Hadassah College Fall 46 Congestion Control Queuing theory Congestion Control Buffer throughput Assumptions Segments arrive independently (Poisson statistics) Random length (bytes) Average arrival rate in steady state Segments leave independently (Poisson statistics) Average emptying rate in steady state (Over)-simplified throughput model receive rate throughtput = maximum receive rate arrival rate buffer utilization = empty rate latency throughput at receivers buffer utilization (from all senders) Results arrival rate ρ = Utilization = empty rate Latency = = empty rate arrival rate empty rate ρ ρ Buffer Level = Latency arrival rate = ρ latency buffer level Utilization ρ Realistic throughput behavior High arrival rate at buffer Longer latency + overflow timeouts latency Re-transmit more segments higher arrival rate at buffer throughput at receivers buffer utilization (from all senders) Computer Networks Hadassah College Fall 47 Computer Networks Hadassah College Fall 48

13 49 TCP Congestion Control End-to-end congestion control Based on host estimates No feedback from intermediate network nodes Slow-start Begin session with low transmission rate Increase rate until timeouts begin Fast retransmit Do not wait for timeout Re-transmit after duplicate s Congestion avoidance Limit transmission rate after duplicate s Transmission rate initial slow-start rate Fast recovery Congestion avoidance with larger transmission rate Slow Start Congestion window (cwnd) Source window Maximum number of "uned" bytes Initial cwnd = MSS (maximum segment size) Data rate = MSS / RTT Exponential growth On each cwnd cwnd + size of data ed if (cwnd > maximum cwnd = destination window) cwnd max cwnd if ( timeout) Segment size threshold (ssthresh) last cwnd cwnd initial cwnd = MSS RTT MSS MSS MSS Computer Networks Hadassah College Fall Computer Networks Hadassah College Fall 5 Fast Retransmit Better performance with long timeout timer duplicate s for segment re-send segment for missing SEQ_ SEQ_ SEQ_ SEQ_4 SEQ_5 (duplicate) _ (duplicate) _ (duplicate) SEQ_ (duplicate) _5 Congestion Avoidance Fast retransmit options Tahoe protocol duplicate s timeout cwnd initial cwnd = MSS Reno protocol Re-transmit segment cwnd initial cwnd = MSS duplicate s Fast re-transmit cwnd cwnd / On each, cwnd cwnd + MSS (exponential growth linear) Fast recovery (Reno option) duplicate s Fast re-transmit cwnd ssthresh Computer Networks Hadassah College Fall 5 Computer Networks Hadassah College Fall 5