Network Performance Management

1
Network Performance Management

• S. Keshav
• C/NRG
• (with Rosen Sharma, Andy Choi, Wilson Huang, Lili
  Qiu, Russell Schwager, Rachit Siamwalla, Jia
  Wang, and Yin Zhang)

2
Motivation

• Networks are increasing in breadth.
• greater density of connections
• PCs come with built-in networking
• ADSL and cable modems
• wireless networking
• as well as in depth
• variety of qualities, policies, and media

3
The current situation

• Loss of productivity from
• slow file access
• web site disconnection
• slow access to a web site
• no one knows exactly why!
• Greater breadth and depth gt even more
  dependency on the network gt even more problems

4
Is QoS enough?

• Lots of research in the area of QoS
• RSVP, differential service etc. provide a good
  overall user experience, one stream at a time
• Is QoS all there is to a good user experience?
• An incorrect reservation ? poor service for one
  stream
• A misconfigured router ? complete loss of service
  to one or more ports!

5
Aha!

• User experience is affected more by mundane
  network management than by exotic QoS research
• This motivates our entire research effort

6
Why networks fail

• Link or router failure
• Transient overload
• Unanticipated increase in load
• Misconfiguration

Increasingly harder to detect
7
Need Better Network Management

• Current approaches
• GUI-centric
• lots of flashing lights, but no intelligence
• Can detect failures but...
• ad hoc capacity planning
• ad hoc configuration
• no way of testing other than just try it!
• Cant manage network performance

8
Performance management
Topology discovery
Configure new hardware (simulation)
Collect statistics (monitoring)
Fix problems (AI and simulation)
Identify problems (display and simulation)
9
Discovery Project Octopus
Temporary Set
Heuristic
Permanent Set
10
Techniques

• DNS-ls
• SNMP
• Random probe
• Traceroute
• Directed broadcast ping

11
Results

• Have automatically discovered entire CS
  department topology
• As well as entire Stanford topology (gt 220
  subnets)
• Cornell topology is being discovered as we speak!
• info being shared with CIT

12
Monitoring

• A PERL script uses SNMP and queries a router
  using various MIB entries.
• The MIB entries are stored in an input file.
• The values gathered from the router are stored in
  a file.
• The script works on both UNIX and WinNT.

13
Monitoring (contd.)

• Other PERL scripts parse the data and convert it
  to other formats.
• Currently supported formats
• HTML - The data is presented in a table format in
  HTML.
• GNUPlot graphs - The data can be graphed or saved
  in pbm format

14
A Case Study CSGate2

• From 2/19/98 to 2/23/98, the router CSGate2 was
  probed every 5 minutes recording various
  statistics on the data coming into and going out
  of the router.

Incoming bytes at CSgate2
15
Display goals

• We want to display multiple views
• Views should be dynamic
• Shoul allow expansion and contraction
• Rapid creation of user interface
• Reusability of GUI components

16
Solution Script Java

• Component-based system
• Reusable manageable components
• Can build large manageable applications
• Sharing over the web
• Record and playback

17
Architecture

• Use JavaScript/Visual Basic as the scripting
  language
• Use Java to write components
• Create a adapter hierarchy for the current AWT
  components

18
Script Java

• Objects
• HTML pages
• Java structures intelligence
• protection by namespace

• Data Model
• linearized data structures
• java ? perl ? javascript

• Communication Abstraction
• multicast channels

19
Advantages

• Allows us to glue components using a scripting
  language, allowing rapid prototyping and
  development
• New components can be easily integrated
• For large applications, a lot of the complexity
  and chaos can be taken out of scripting

20
Advantages(cont.)

• JavaScript can be streamed from the server,
  allowing for presentations and sharing
• Dynamic Html
• layers are windows
• these windows render html

21
Storage goals

• We need to store topology and monitoring results
  somewhere
• Database too structured and too much overhead
• File system not enough semantics
• Idea treat URL as a file system link and HTML
  tags as associated semantics

22
WebFS

• HTML tags allow arbitrary semantic abstractions
• Manipulate these abstractions to present a
  virtualized file system
• grep -headings .html
• sed /ltannot tagfoogt/jdbc(tags.db, foo)/

23
The magic bullet simulation

• Realistic simulation where networking subsystem
  interacts with other parts of kernel
• Fast simulation for large networks ( gt 1000
  hosts)
• Hide the abstraction of simulated network, same
  API as system calls

24
FreeBSD kernel ?User Space
gated
traps
?
Telnetd
ping
Sockets
Network Stack
25
Simulated machine

• Task based approach
• a trap sends a message to kernel
• an upper call is a message from kernel
• All components of simulated machine live on same
  process

Simulated link
26
More on simulated machine

• Capture network related system calls, file
  descriptor auto re-mapping.
• Virtual file system root
• Single-thread kernel, therefore no need for
  locking

27
Simulated network
machine
gated
msg
Telnetd
ping
Kernel core
28
Integrating with real network

• Use U-Net to interact with external device
• Router has the illusion of being in a physical
  network
• Test equipment before actual deployment

Unet
Physical Router
29
Tradeoffs

• Balance between realism and speed
• Using FreeBSD as basis for realistic simulation
• Using session level simulation to speed up
• Ease of porting applications

30
Open issues

• Fault identification
• Bayesian networks?
• Ensemble of experts?
• Other AI approaches?
• How to do session-level simulation?
• Configuring real systems
• IP9000