Transcript LAN design issues

LANS, performance and
Client/Server design issues
Basic performance definitions
Bandwidth

Raw data rate of links
Capacity


Theoretical limit of data transfer
Measured over the network, sub-net or link
Throughput


Actual data transmitted (e.g. packets per second)
Limited by protocol overhead, delays, latency etc
Throughput v Capacity
Optimum
100%
Max throughput
Throughput
Max capacity
0%
Actual
Load
Basic performance definitions
Latency

End-to-end delay, comprising
 propagation delay (near speed of light),
 transmission delay (media speed),
 store-and-forward delay (bridge/switch/router buffering),
 processing delay (action on protocol elements)

Sensitivity to delay is application dependent
 video is very sensitive and
 virtual terminal (Telnet) is medium sensitive (user-
dependent)
Basic performance definitions
Jitter


The variability of latency
Buffering can smooth the delay
Media access delay

LAN access delay depends on
 Access scheme used
 No. of contending devices
Accuracy



Data corruption
Bit error rate on WAN links
< 1 in 106 on LANs
Key performance relationships
Payload (TCP/IP over Ethernet)



Payload = MTU – (TCPOverhead + IPOverhead+ MACOverhead)
MTU is maximum transmission unit
Overheads are: TCP 20 bytes; IP 20 Bytes; MAC 18 bytes
Maximum packet rate
 PPSmax =Channel Speed
(8 bits x PDUsize )
For example at 64 kbps with 128 byte PDUs
PPSmax =64000/(8 x 128) = 62.5 pps
Performance issues
Different network types have different
maximum packet/frame sizes
Overlarge packets need fragmentation
and re-assembly to be transmitted


limits throughput
reduces performance
Compression can be used to improve
performance on slower speed links
Key performance relationships
Packet rate and link speed

Ensure links do not exceed PPSmax
Error probability and frame size



Larger packets are more likely to contain an error
Protocol efficiency E
E=
Sdata
_
[R(Sdata+Sprot+Sack)]

Sdata= data size; Sprot=protocol overhead; Sack = ack size

R = expected number of transmissions per packet

Or R=1+packet error rate e.g 1.001 if 1 in 1000 errors
Typical bottlenecks
Shared services (centralised servers etc)
Multi-user applications and databases
Low-speed NICs
Shared LAN segments
Low-bandwidth WAN links
Core routing and switching components
Firewalls (particularly public-facing)
Inappropriate compression usage
Main types of server
File Servers
Database Servers
Transaction Servers
GroupWare Servers
Web Servers
Middleware
Resides between the client and server
Gives the single system image
Typically a major component in a NOS
Provides: directory services, network
security etc
Contains proprietary elements where
required
Scalable Client Server
For the single User

Client, middleware and most of the business
services on a single machine
For the SME


Use of small LAN
Often involves multiple clients talking to a local
server
For the Enterprise


Connection of multiple servers across a network
To utilise fully requires low cost, high speed
bandwidth
Features of Server S/W
Wait for client initiated requests
Execute many requests at the same time
Are able to prioritise requests
Can run activities in background
Are resilient and keep running
Main contenders;



Netware
Windows (and NT) Server
Unix/Linux
Features of Client S/W
Communicate service requests to a server
Needs to be robust
Provide protection from programs that crash
Provide a mechanism for file transfer
Provide multi tasking
Allow background processes to take place
Client/Server bottlenecks
Client and servers are subject to
constraints from




Memory
CPU cycles
Network and disc input/output
System bus throughput
Client/Server Design Issues
User requirements (applications, response
rate, latency etc)
NOS (free choice or pre-determined)
Topology (technology determined)
Server placement (on the network)
Thick/thin client (balance of services)
Groupware (CSCW) use
Maintenance (ability/cost)
Protocol Issues
TCP/IP protocol performance depends on







The implementation/stack used
The OS and platform
Packet size distribution of the application
Background traffic characteristics of the contended
paths
LAN, MAN, WAN media properties , overheads and
BERs
Intermediate device-forwarding characteristics
TCPs sliding window behaviour
Typical bottlenecks
The LAN/WAN interface

WANs are typically an order of magnitude slower
Routers need to buffer WAN traffic


Buffers require sufficient memory
Insufficient buffer space leads to more retransmissions – lowering efficiency
Queuing/buffering also increases end-to-end
latency

Some applications may not tolerate high latency,
timeout and re-transmissions will occur increasing
the problem
Data modelling
Gather information of the users to derive

Application maps
 Which are used and where

Data flow
 How much data flows from machine to machine
 Traffic types




Terminal/host, Client/Server, Peer-to-peer, Server-to
server, Distributed entity traffic
Local:Remote 80:20 50:50 in modern intranets
Build user-type and server profiles
Traffic matrices
 Characterise data in and data out of each site
Hierarchical network design
Three-layer architecture

Backbone layer
 High-speed switching layer
 Mesh design for resilience/minimise outages

Distribution layer
 Link points between campus LANs and core backbone

Access layer
 End user interface
 Typically LAN environment
Advantages of hierarchical
network design
Scalability

Easier to add to the network
Manageability

Easier to identify location of problems
Broadcast traffic segmentation


Traffic confined to smaller broadcast
domains
Less traffic over expensive links
Ethernets
Generic Ethernet design rules

Max. stations in a collision domain =1024







(collision domain is where the time taken to transmit a min. frame
is shorter than the time to detect a collision)
Only use repeaters at link-ends
Avoid exceeding standard specs
No more than 4 repeaters in a collision domain
No more than 3 coax segments in a collision domain
Inter-repeater links are best implemented by fibre
(10baseFL, 10baseFB) or 10baseT
10base5, 10base2 and 10baseT can be mixed if wanted
LAN performance considerations
Fixed parameters

Bit rate, slot time etc
Variable factors






Packet length distribution
No.of hosts in a collision domain
Arrival rate of frames
Average length of cable
Distance between nodes
Average medium acquisition time
Ethernet design rules
To optimise performance




Use shorter cables - Long cables increase
collision detection time
Do not attach too many nodes to a
segment
Use largest possible packet size – this
reduces collisions
Try not to mix real-time and heavy bulk
data traffic in the same collision domain
VLANs
Logical hierarchy imposed on a flat
switched network allowing




Scalability
Formation of workgroups
Simplified admin
Better security
Wireless LANs
Use Wireless LAN access points(WLAP)

Simplest LAN use single WLAP
 Effectively a wireless star topology

Multiple WLAPs can be used
 Can incorporate wired and wireless segments
WLAPS can support
 10-50 clients
 Over a 30-60m radius (depends on radio transmission
environment)
Wireless LANs can simplify installation and reduce
costs – especially in smaller and older buildings
Summary
Good design should optimise
performance
Many factors affect performance



Technology
Software tuning
Physical environment
The interaction of all network
components needs to be considered