Network Protocols Sarah Diesburg Operating Systems CS 3430

Network Protocols Sarah Diesburg Operating Systems CS 3430

Distributed Systems Allow physically separate computers to work together + Easier and cheaper to mass-produce simple computers Off-the-shelf components + A company can incrementally increase the computing power

Promises of Distributed Systems Higher availability Better reliability If one machine goes down, use another A user is able to store data in multiple locations More security Each simple component is easier to make secure

Reality of Distributed Systems Worse availability Worse reliability One can lose data if any machine crashes Worse security A system may depend on many or all machines being up Security is as strong as the weakest component Coordination is difficult because machines can only use the network medium

Network Technologies Definitions Network: physical connection that allows two computers to communicate Packet: a unit of transfer A sequence of bits carried over the network Protocol: An agreement between two parties as to how information is to be transmitted

Broadcast Networks A broadcast network uses a shared communication medium e. g. wireless, Ethernet, cellular phone network The sender needs to specify the destination in the packet header So the receiver knows which packet to receive If a machine were not the intended destination Discard the packet

Arbitration Concerns the way to share a given resource In Aloha network (1970 s) Packets were sent through radios on Hawaiian Islands

Aloha Network Arbitration: blind broadcast, with a checksum at the end of a packet Packets might become garbled in the case of simultaneous transmissions

Aloha Network Arbitration: blind broadcast, with a checksum at the end of a packet Packets might become garbled in the case of simultaneous transmissions

Aloha Network Arbitration: blind broadcast, with a checksum at the end of a packet Packets might become garbled in the case of simultaneous transmissions

Blind Broadcast Receiver: If a packet is garbled discard else sends an acknowledgement Sender: If the acknowledgement does not arrive resend the packet

Ethernet (introduced in the early ‘ 80 s) By Xerox First practical local area network Uses wire (as opposed to radio) Broadcast network Key advance: a new way for arbitration

Ethernet’s Arbitration Techniques Carrier sensing: Ethernet does not send unless the network is idle Collision detection: sender checks if packet is trampled If so, abort, wait, and retry Adaptive randomized waiting: a sender picks a bigger wait time (plus some random duration) after a collision

The Internet A generalization of interconnected local area networks Uses machines to interconnect various networks Routers, gateways, bridges, repeaters LAN 2 Act like switches Packets are copied as they transmitted across different networks LAN 1

Routing Concerns how a packet can reach its destination Typically, a packet has to go through multiple hops before getting to a destination Each hop is a router, which directs a packet to the next hop Routing is achieved through routing tables

Routing Table Updates Each routing entry contains a cost 1. 2. 3. <destination, next hop, # hops> Neighbors periodically exchange routing table entries If the neighbor has a cheaper route, use that one instead

Point-to-Point Networks Instead of sharing a common network medium, all nodes in the network can be connected directly to a router/switch

Point-to-Point Networks + Higher link performance (no collisions) + Greater aggregate bandwidth than a single link

Point-to-Point Networks + Network capacity can be upgraded incrementally + Lower latency (no arbitration)

Issues in Point-to-Point Networks Congestion occurs when everyone sends to the same output link on a switch Crossbar buffers

Networking: Physical Reality vs. Abstraction Physical reality: packets Abstraction: messages Limited size Arbitrary size Unordered Ordered Unreliable Reliable Machine-to-machine Process-to-process Only on local area network Routed anywhere Asynchronous Synchronous Insecure Secure

Arbitrary-Size Messages Can be built on top of limited-size ones By splitting a message into fix-sized packets Checksum can be computed on each fragment or the whole message

Internet Protocol (IP) Provides unreliable, unordered, machine-tomachine transmission of arbitrary-size messages

Process-to-Process Communications Built on top of machine-to-machine communications through the use of port addresses Each message contains the destination port to talk to the correct process

Unreliable Data Protocol (UDP) Provides unreliable, unordered, user-to-user communication Built on the top of IP

Ordered Messages Built on top of unordered ones Use sequence numbers to indicate the order of arrival Specific to a connection If packet 3 arrives before packet 2, wait for packet 2. Always deliver packets in order, to user applications

Reliable Message Delivery Built on top of unreliable delivery Problem: Network infrastructure can garble messages Packets can be dropped if network buffers are full

Solution Checksum each message At a receiver, discard messages with mismatching checksums A receiver acknowledges if a packet is received properly A sender resends the same message after not hearing the acknowledgment for some time (a timeout period)

A Minor Problem A sender may send twice, if the first acknowledge is lost The receiver needs to discard duplicate packets

Implications A sender needs to buffer messages that are not yet acknowledged The receiver must track messages that could be duplicates

Transmission Control Protocol (TCP) Provides a reliable byte stream between two processes on different machines over the Internet sequence number: 1 checksum: fa 73 cd 10

Transmission Control Protocol Fragments the byte stream into packets and hands them to IP

TCP Message Categories Sender Sent and acknowledged Sent and not acknowledged Not yet sent Receiver Forwarded to application Received and buffered Not yet received

More on the Sequence Number Need a way to recycle sequence numbers Each TCP packet has a time-to-live field If the packet is not delivered in X seconds The packet is dropped Sequence numbers can be reused An epoch number used to identify which set of sequence numbers is being used Incremented at each boot Stored on disk

Congestion Implications of timeout period at a sender Too long unnecessary waiting Too short a message is transmitted when an acknowledgement is in transit Network congestion delayed acknowledgement timeout data retransmission more congestion

TCP Solution Slow start: TCP starts by sending a small amount of data If no timeout, more data is sent If timeout, TCP reduces the amount of data being sent

Distributed Transaction Multiple machines agree to do something atomically, but not necessarily at exactly the same time Mechanism: two-phase commit

Two-Phase Commit Account X Account Y Phase 1: ask if each can commit 1. Begin transaction Ask Y for $1 Enough cash 2. Write “Y = Y - $1” Ready to commit Phase 2: commit 3. Write “X = X + $1” 4. Commit Ask Y to commit 5. Commit

Scenarios If X crashes between 1 and 2 If X crashes before step 4 Y will wake up and do nothing X will timeout and abort the transaction X will wake up and abort the transaction If X crashes between 4 and 5 Y will timeout and ask X for the transaction

Scenarios If Y crashes between 2 and 5 Y will wake up and check the log When X sends Y the commit message, Y will commit Y can also timeout and ask X the current status