1. Introduction Outline Statistical Multiplexing Network Architecture
Performance Metrics (just a little)
Spring 2006
CS 332

2. Our Journey Peterson Text: Chapter 1
Chapters 4,5,6,8 (with Chapters 2,3,7 sprinkled in as needed)
Why?
Spring 2006
CS 332

3. Building Blocks … Nodes: PC, special-purpose hardware…
hosts
Switches (host connected to at least two links that runs software that forwards data received on one link out on another).
Links: coax cable, optical fiber…
point-to-point
multiple access
Be sure to mention media access protocols when you have multiple access (maybe talk a little about Ethernet and it’s media access protocol.
Note difference between switch, host, and router (and that host can also be router). …
Spring 2006
CS 332

4. Switched Networks A network can be defined recursively as...
two or more nodes connected by a link, or
two or more networks connected by two or more nodes
Nodes on inside of
cloud called switches,
nodes on outside
called hosts
internetwork
Mention that single host can obviously have more than one network interface.
Cloud can Denote any
Type of network:
P2P, multi-access,
switched
Node between networks
called router or gateway
Spring 2006
CS 332

5. Routing
Just because we have physical connectivity between all hosts, doesn’t mean we have provided host-to-host connectivity
Need to be able to specify hosts with which we wish to communicate
I.e. each host needs an address (note that the address is really only a name – it need not relate in any way to the physical location of the host)
Routers and switches between hosts need to use address to decide how to forward messages
Routing: process of determining systematically how to forward messages toward the destination node based on the destination node’s address
Spring 2006
CS 332

6. Not Just One Destination…
Unicast – single source, single destination
Broadcast – single source, all destinations (well, sort of)
Multicast – single source, whole group of destinations
Anycast?!
Why would you want this?
Spring 2006
CS 332

7. A Key Idea
We can define a network recursively as a network of networks.
At bottom layer, it is implemented by some physical medium
At higher layers it is a “logical” network
Key issue: how do we assign addresses at each layer in a manner which allows us to efficiently route messages?
Spring 2006
CS 332

8. Strategies Circuit switching: carry bit streams
original telephone network
Connection is established before any data sent
Packet switching: store-and-forward messages
Internet (Why?)
Send discrete blocks of data from node to node (called a packet or message)
Why packet switching versus circuit switching? Efficiency for data traffic. Also reliability in face of downed link (recall that Internet was a military network, sort of like Interstate system).
Note that these can be logical things too (as we’ll see).
Spring 2006
CS 332

9. Virtual Circuit Switching
Explicit connection setup (and tear-down) phase
Subsequence packets follow same circuit
Sometimes called connection-oriented model 1 3 2 5 11 4 7
Switch 3
Host B
Switch 2
Host A
Switch 1
Analogy: phone call
Each switch maintains a VC table
Advantages: mechanics of forwarding packets are easier (and thus quicker)
Disadvantages: Explicit setup/teardown. What happens if link goes down then
Comes back up!
Advantages?
Disadvantages?
Spring 2006
CS 332

10. Datagram Switching No connection setup phase
Each packet forwarded independently
Sometimes called connectionless model 1 3 2
Switch 3
Host B
Switch 2
Host A
Switch 1
Host C
Host D
Host E
Host F
Host G
Host H
Analogy: postal system
Each switch maintains a forwarding (routing) table
Spring 2006
CS 332

11. Multiplexing Time-Division Multiplexing (TDM)
Frequency-Division Multiplexing (FDM) L1 L2 L3 R1 R2 R3
Switch 1
Switch 2
STDM: time broken into cycles, each flow gets to transmit during its cycle.
FDM: each flow gets a given frequency band, each flow transmits only in its band.
Problem: wasted resources: Nodes may wait for no one (think traffic light) in STDM, and same with FDM, since a given frequency may be free. (So frequency hopping can be used). This is done statistically to try to minimize “collisions”.
Also, for both, need to know max number of flows ahead of time.
Note that in either
technique, bandwidth
can be wasted!
Spring 2006
CS 332

12. Statistical Multiplexing
On-demand time-division (rather than in specific time slot)
Schedule link on a per-packet basis
Packets from different sources interleaved on link
Buffer packets that are contending for the link
Buffer (queue) overflow is called congestion
Fairly allocating
link capacity and
dealing with congestion
are key issues here!
May want to mention things like head-of-line blocking and queue placement in switches. …
Spring 2006
CS 332

13. IPC Abstractions Stream-Based Request/Reply video: sequence of frames
video applications
on-demand video
video conferencing
Request/Reply
distributed file systems
digital libraries (web)
Request/Reply is really client/server type. Properties you might want include RELIABLE MESSAGE DELIVERY (THEY ALL DO ARRIVE), ONLY ONE COPY RECEIVED, MIGHT PROTECT PRIVACY AND DATA INTEGRITY SO NO EVESDROPPING
Stream based MIGHT NOT NEED DELIVERY GUARANTEE (VIDEO IS FINE WITH A FEW MISSING FRAMES) BUT WOULD PROBABLY REQUIRE PACKETS MAINTAIN SEND ORDER, MIGHT BE PARAMETRIZED TO SUPPORT DIFFERENT DELAY PROPERTIES OR ONE-WAY OR TWO-WAY STREAMS.
Spring 2006
CS 332

14. Host-to-host connectivity
Layering
Use abstractions to hide complexity
Abstraction naturally lead to layering
Alternative abstractions at each layer
Application programs
Request/reply
Message stream
channel
channel
Host-to-host connectivity
Hardware
Spring 2006
CS 332

15. Protocols Building blocks of a network architecture
Each protocol object has two different interfaces
service interface: operations on this protocol
peer-to-peer interface: messages exchanged with peer
Term “protocol” is overloaded
specification of peer-to-peer interface
module that implements this interface
Spring 2006
CS 332

16. Interfaces Spring 2006 CS 332 Host 1 Host 2 Service High-level
object
interface
object
Protocol
Protocol
Peer-to-peer
interface
Spring 2006
CS 332

17. Protocol Machinery Protocol Graph
most peer-to-peer communication is indirect
peer-to-peer is direct only at hardware level
Host 1
Host 2
File
Digital
Video
Digital
File
Video
library
library
application
application
application
application
application
application
RRP
MSP
RRP
MSP
HHP
HHP
RRP: request/reply protocol
MSP: msg stream protocol
Spring 2006

18. Machinery (cont) Multiplexing and Demultiplexing (demux key)
Encapsulation (header/body)
Host 1
Host 2
Application
Application
program
program
Data
Data
RRP
RRP
RRP
Data
RRP
Data
HHP
HHP
HHP
RRP
Data
Spring 2006
CS 332

19. Internet Architecture
Defined by Internet Engineering Task Force (IETF)
Hourglass Design
Application vs Application Protocol (FTP, HTTP) …
FTP
HTTP NV
TFTP
TCP
UDP IP
NET 1 2 n
Spring 2006
CS 332

20. ISO Architecture Note transport layer is “end-to-end” Spring 2006
End host
End host
Note transport
layer is “end-to-end”
Application
Application
Presentation
Presentation
Session
Session
Transport
Transport
Network
Network
Network
Network
Note a key difference between UDP and IP. Also between network layer and transport layer.
Talk here also about the user/kernel space boundary, about network interface cards, and perhaps about the path that a message takes as it goes from a user process out onto the network. Might mention a bit of architecture, and why the memory bandwidth can have a big impact on network performance (and how each copy operation limits overall performance). Increasing processor speeds but slowly increasing memory speeds means copying a problem. See p. 51 in text “copying data from one buffer to another is one of the most expensive things a protocol implementation can do!”
Talk about the problems with this in terms of the changing relation between processor and link speeds. That is, in the old days, the link was the bottleneck, so a big protocol stack wasn’t an issue. Now the problem is the protocol stack.
Data link
Data link
Data link
Data link
Physical
Physical
Physical
Physical
One or more nodes
within the network
Spring 2006
CS 332

21. Performance Metrics Bandwidth (throughput) “width” of pipe
data transmitted per time unit
link versus end-to-end
notation
KB = 210 bytes
Mbps = 106 bits per second
Latency (delay) “length” of pipe
time to send message from point A to point B
OR time for bit to travel from point A to point B
Also “link latency”
one-way versus round-trip time (RTT)
components
Latency = Propagation + Transmit + Queue
Propagation = Distance / c
Transmit = Size / Bandwidth
Spring 2006
CS 332

22. Bandwidth versus Latency
Relative importance
Infinite bandwidth
Spring 2006
CS 332

23. Protocol Implementation Issues
Which process model?
Process-per-protocol model
Each protocol implemented by separate process (thread)
Sometimes logically “cleaner” – one protocol, one process
Context switch required at each level of protocol graph
Process-per-message model
Each message handled by a single process with each protocol a separate procedure that invokes the subsequent protocol
Order of magnitude more efficient (procedure call much more efficient than context switch)
Spring 2006
CS 332

24. Protocol Implementation Issues
Service Interface relationship with process model
If high-level protocol invokes send() on lower level protocol, it has message in hand so no big deal
If high-level protocol invokes receive() on lower level protocol, it must wait for receipt of message, which basically forces a context switch.
No big deal if app directly calls network subsystem, but big deal if it occurs at each layer of protocol stack
Cure: low level protocol does an upcall (a procedure call up the stack) to deliver message to higher level
Spring 2006
CS 332

25. Protocol Implementation Issues
Message Buffers
In socket API, application process provides buffers for both outbound and incoming messages. This forces top most protocol to copy messages from/to network buffers.
Copying data from one buffer to another is one of the most expensive operations a protocol implementation can perform.
Memory is not getting fast as quickly as processors are
Solution: Rather than copying from buffer to buffer at each layer of protocol stack, network subsystem defines a message abstraction shared by all protocols in the protocol graph.
Can be viewed as string manipulations with pointers
Note: you can’t move data any faster than the slowest copy operation!
Spring 2006
CS 332

26. Implementation
Are you using streams or request/reply, and what are the ramifications?
What operating system are you coding on/for and where is the sockets library?
What languages can you use, in theory?
Why would you wish to use specific languages?
What will you have to do in your first assignment?
Spring 2006
CS 332

27. Implementation Port Numbers Endian issues Compiler flags
Solaris: reserved ( ), ephemeral( )
BSD:reserved(1-1023), ephemeral( ), nonprivileged servers( )
IANA: well known (1-1023), registered( ), dynamic ( )
Endian issues
Compiler flags
Solaris –lsocket –lnsl
Linux no flags required
Specifying command line arguments
Null characters in strings?
Spring 2006
CS 332