Flexlab: A Realistic, Controlled, and Friendly Environment for Evaluating Networked Systems

1
Flexlab A Realistic, Controlled, and Friendly
Environmentfor Evaluating Networked Systems

• PRESENTED BY AYMAN EL-SHAYEB

2
Objectives

• Understanding the Flexlab system
• Understand the deployment process
• Understanding the Emulab system
• Understanding the Planet-lab system
• Understanding the datapository
• Understanding the network measurements

3
Agenda

• Network System development Life Cycle
• Emulab overview
• Planet-Lab overview
• Datapository
• FlexLab system
• Reverse- engineering the internet

4
Network System development Life Cycle

• PLANETLAB (Princeton University PlanetLab
  Consortium)

5
Ideas Maturity

• PLANETLAB (Princeton University PlanetLab
  Consortium)

6
Emulab System

• Emulab is a network testbed, giving researchers a
  wide range of environments in which to develop,
  debug, and evaluate their system

EmuLab site
7
Steps to make Emulab project

• Logging into web interface
• Design network topology
• The topology can be descried by GUI or NS
  (Network simulation) format
• bandwidth, delay and loss can be specified in
  each link
• Operating system Image of deferent Operating
  system can be specified for each PC node
• Choose begin experiments from web interface
• An email will be sent by status (success or fail)
• In case of success an email is sent by topology
  details IPS , router tables and operating
  systems
• Log by Secure shell ( SSH) in any nodes. Install
  the application and run

• EmuLab site

8
Example to topology

• EmuLab site

9
Planet-Lab

• Planet-Lab is a global overlay network for
  developing and accessing broad-coverage network
  services.
• It includes a large set of geographically
  distributed machines
• It is a realistic network so it includes
  congestion, failures and link behaviors

PlanetLab An Overlay Testbed for Broad Coverage
Services
10
Planet-Lab

• 784 servers at over 382 sites

11
Planet-Lab

• The center piece of the Planet-Lab architecture
  is a slice (a horizontal cut of global Planet-Lab
  resources)
• Each service (a set of distributed and
  cooperating programs delivering some higher-level
  functionality) runs in a slice of Planet-Lab.

PlanetLab An Overlay Testbed for Broad Coverage
Services
12
Planet-Lab

• A slice is a network of virtual machines, with a
  set of local resources bound to each virtual
  machine.
• A virtual machine is the environment where the
  program that implements some aspect of a service
  runs.
• Each virtual machine runs on a single node and is
  allowed to consume some fraction of that nodes
  resources.

PlanetLab An Overlay Testbed for Broad Coverage
Services
13
PlanetLab
PlanetLab An Overlay Testbed for Broad Coverage
Services
14
Planet-Lab

• A node manager, implemented as part of the VMM,
  runs on each node. It takes a set of tickets as
  input.
• If the request can be satisfied, the node manager
  reserves the specified resources, creates a
  virtual machine, binds it to those resources, and
  returns a lease.
• This lease is later used by the service manager
  to launch a program within the virtual machine.

PlanetLab An Overlay Testbed for Broad Coverage
Services
15
Planet-Lab

• A resource monitor is an service running on each
  node. It monitors resource availability on the
  node and periodically reports the results to one
  or more agents.
• An agent collects resource availability
  information from a set of resource monitors or
  other agents, and issues tickets that can be used
  to acquire resources on the monitored nodes.
• An agent can respond to two kinds of queries
• It can advertise the tickets it is holding that
  meet certain criteria
• It can grant tickets themselves to a requester.
  The agent may also overbook the resources
  available on the nodes it is responsible for

PlanetLab An Overlay Testbed for Broad Coverage
Services
16
Planet-Lab

• A resource broker responds to queries from
  service managers trying to discover a slice to
  run in.
• Each query describes the resources needed by the
  service, and specifies the principal on whose
  behalf the request is being made.
• Given a query, the broker first contacts one or
  more agents to discover what tickets it is
  possible to obtain. It then matches this
  information with the services resource
  requirements to produce a slice specification,
  contacts the relevant agents to obtain the
  tickets, and returns them to the service manager.

PlanetLab An Overlay Testbed for Broad Coverage
Services
17
Planet-Lab

• A service manager is associated with each
  service. It contacts a resource broker to
  discover a slice and obtain the tickets needed to
  instantiate it.
• It then submits the tickets to the admission
  control mechanism on each node to create a
  network of virtual machines.
• Should virtual machine creation succeed on each
  node, the service manager then launches the
  service, that is, loads and starts a program in
  each virtual machine.
• Admission control returns a lease on the slice.
  The manager must periodically renew this lease.

PlanetLab An Overlay Testbed for Broad Coverage
Services
18
Planet-Lab drawback

• Scheduling Accuracy
• PlanetLab is heavily overloaded so it causes
  significant accuracy problems
• Network Conditions
• It can change frequently at small time scales

• FlexLab A Realistic, Controlled , and Friendly
  Environment

19
Challenges and Opportunities in Internet Data
Mining

• Communication networks produce vast amount of
  data
• On short timescales, the data can facilitate
  network monitoring and troubleshooting
• Over minutes to days, the data is useful for
  network traffic engineering.
• Over months, the data is useful for capacity
  planning
• Data can be used in simulation and emulation

• Challenges and Opportunities in Internet Data
  Mining

20
Purpose of paper

• Proposed standard way to collecting and archiving
  large volume of network measurement data
  (datapository)
• Examines several challenges and opportunity in
  collecting and archiving data

• Challenges and Opportunities in Internet Data
  Mining

21
Challenges

• Robust, distributed data collection
• Multi-timescale , heterogeneous data
• Non-standard, irregular data
• Large archival data mining
• Workflow metadata and privacy

• Challenges and Opportunities in Internet Data
  Mining

22
The datapository

• Challenges and Opportunities in Internet Data
  Mining

23
The datapository

• Data Collection
• Data types
• Data formats
• Access methods
• Feed
• Workflow

• Challenges and Opportunities in Internet Data
  Mining

24
The datapository

• Data storage
• MySQL engine to use SQL interface
• Analysis and queries
• Network operator
• Researchers performing analysis
• Simulation/emulation system

• Challenges and Opportunities in Internet Data
  Mining

25
The datapository

• External Interfaces
• Supports XMLRPC export mechanism
• It can export data in various format, including
  human-readable output and Matlab-friendly

• Challenges and Opportunities in Internet Data
  Mining

26
Opportunities

• Repeatable Scientific Experimentation
• Repeating a past experiment
• Performing new analysis
• Network Monitoring and Control
• Trace network changing conditions
• Security threats

• Challenges and Opportunities in Internet Data
  Mining

27
Flex-Lab system

• Flex-Lab is a new testbed environment that
  enables a new generation of network models for
  emulation
• Flex-Lab models end-to-end characteristics of
  Internet paths between pairs of overlay nodes.
• Emulation system disadvantages
• The network condition of emulation system like
  Emulab is artificial.
• The researcher is not sure which network aspects
  are poorly modeled

• FlexLab A Realistic, Controlled , and Friendly
  Environment

28
The Live-network testbeds drawbacks

• It present a challenging environment for
  development, debugging and evaluation
• It has limited resources( CPU, memory, and I/O)
  that are shared among many users and are
  frequently overloaded
• Users cannot perform many privileged operations
  including choosing the OS, controlling network
  stack parameters or modifying the kernel

• FlexLab A Realistic, Controlled , and Friendly
  Environment

29
Overall System Architecture

• Emulab nodes are connected in a full mesh
  abstracting the Internet as a set of pair wise
  network characteristics.
• each node to participate in a Flexlab link is
  associated with a PlanetLab node from which it
  will get its network characteristics.

• FlexLab A Realistic, Controlled , and Friendly
  Environment

30
Overall System Architecture

• The network model feeds network parameters into
  the path emulator
• The network model can also optionally use data
  from the measurement repository
• The model sends network parameters using Emulabs
  event system.

• FlexLab A Realistic, Controlled , and Friendly
  Environment

31
Overall System Architecture

• This framework enables transparent switching
  between different models, and even between
  Flexlab experiments and running live on
  Planet-Lab
• Enables development/debugging under simple,
  predictable models without missing of realistic.

• FlexLab A Realistic, Controlled , and Friendly
  Environment

32
Overall System Architecture

• FlexLab A Realistic, Controlled , and Friendly
  Environment

33
Flexlab models

• Simple static Network Model
• Simple Dynamic Network Model
• Application-Centric Internet Model

• FlexLab A Realistic, Controlled , and Friendly
  Environment

34
Simple static Network Model

• This model set historical data at the beginning
  of the experiment and does not change them
  thereafter
• The historical data is simple static parameters
  such as average latency and bandwidth . It
  changes by low-frequency
• The data can be gotten from archived data in the
  public Datapository

• FlexLab A Realistic, Controlled , and Friendly
  Environment

35
Simple Dynamic Network Model

• This model allows the experimenter to control
  measurement frequency so that path of interest
  can be monitored at higher frequency
• This model uses information from the application
  monitor
• Experimenter can limit the model to certain pairs

• FlexLab A Realistic, Controlled , and Friendly
  Environment

36
Simple Dynamic Network Model

• Further work
• Using archived measurements in the Datapository
  to replay the experimental

• FlexLab A Realistic, Controlled , and Friendly
  Environment

37
Application-Centric Internet Modeling

• It models the internet as perceived by the
  application

• FlexLab A Realistic, Controlled , and Friendly
  Environment

38
Functional scenario

• The model receives characteristics of the
  application's offered load from the application
  monitor
• Then it Replicates that load on Planet-Lab with
  the measurement agent
• the measurement agent determines path
  characteristics through analysis of the TCP
  stream
• Then It sends the results back into the path
  emulator as traffic shaping parameters.

• FlexLab A Realistic, Controlled , and Friendly
  Environment

39
Functional scenario

• FlexLab A Realistic, Controlled , and Friendly
  Environment

40
Design Advantages

• The design creates a feedback loop . Loads and
  emulator are set in near real-time
• The design automatically obtains accurate
  information on how the network reacts to the
  offered load.
• The design allows high-frequently track of
  network changes
• It automatically and quickly detects the
  end-to-end effects of rare events such as outages
  and route flapping, which can be especially
  difficult to model.

• FlexLab A Realistic, Controlled , and Friendly
  Environment

41
Path Emulation

• The emulation uses two queues
• Bandwidth queue (emulates queuing delay)
• Delay queue ( all other source of delay
  propagation , processing, and transmission)
• The most important three parameters
• The size of bandwidth queue
• The drains rate of bandwidth queue
• The time spent in the delay

• FlexLab A Realistic, Controlled , and Friendly
  Environment

42
Parameters Calculation

• Bandwidth size and packet loss
• The author assumes that most packet loss in the
  wired internet caused by congestion
• It is difficult to directly measure the capacity
  of bottleneck router by packets
• Bandwidth queue size is approximated by the
  maximum one way delay
• This solution has difficulty in planet-lab
  because of clock synchronization

• FlexLab A Realistic, Controlled , and Friendly
  Environment

43
Parameters Calculation

• Available Bandwidth
• Some measurement techniques do not take into
  account the reactivity of other flows in the
  network.
• FLexLab measures that flow out into the network
  and measuring the resulting goodput, averaging it
  over the last half second to smooth outliers

• FlexLab A Realistic, Controlled , and Friendly
  Environment

44
Parameters Calculation

• Other Delay
• It records the time each packet is sent, and when
  it receives an ACK for that packet, calculates
  the RTT seen by the most recently acknowledged
  packet.
• Base RTT is calculated as the minimum RTT seen
  recently.
• minimum delay accounts for the propagation,
  processing, and transmission delays along the
  path, with minimum influence from queuing delay.
• The delay queue is set to the Base RTT

• FlexLab A Realistic, Controlled , and Friendly
  Environment

45

• FlexLab A Realistic, Controlled , and Friendly
  Environment

46
Reverse Engineering the Internet

• Reverse engineering is the process of learning
  the design of an object by studying its
  implementation.
• In Internet, design means how its components
  (links, routers, clients, and networks) are
  assembled and configured

• Reverse Engineering the Internet

47
Reverse Engineering the Internet

• There are many efficient network measurement
  techniques that can be used in reverse-
  engineering process like
• tomography can be used to measure link loss.
• tailgating approaches measure link band width
  using few packets

• Reverse Engineering the Internet

48
Reverse Engineering the Internet

• The challenge is to use these techniques to
  produce new measurement tools that are efficient
  enough to run quickly at the scale of the Internet

• Reverse Engineering the Internet

49
What can be measured and analyzed today

• Measured Node Properties
• IP aliases
• Geography
• Owner
• Router role
• Implementation feature

• Reverse Engineering the Internet

50
What can be measured and analyzed today

• Measured Link Property
• Loss
• Reordering
• Delay
• Delay variation
• Capacity

• Reverse Engineering the Internet

51
What can be measured and analyzed today

• Measured Topology Properties
• Topology
• Routing
• Workload

• Reverse Engineering the Internet

52
New Properties

• Failure and Evolution
• Utilization and workload
• Below IP
• Link Pathologies
• Client Location

• Reverse Engineering the Internet

53
Reverse Engineering Cost
54
Reverse Engineering Cost

• Overall cost 5 billion packets , 803 GB
• Per source 25 million packets
• Per source 37 per month per source
• Assuming 200 source or vantage points
• 1 Mbs 100 per month

• Reverse Engineering the Internet

55
Efficient Measurement

• The goal is to make reverse-engineering clearly
  practical and acceptable by reducing this work
  load
• Existing techniques were typically not designed
  to be efficient for internet mapping. We can
  reduce the cost by eliminating the redundancy

• Reverse Engineering the Internet

56
Efficient Measurement

• By eliminating redundancy author saves 5.8
  million packets

• Reverse Engineering the Internet

57
References

• PLANETLAB (Princeton University PlanetLab
  Consortium)
• http//coblitz.codeen.org/roads.cs.princeton.edu/w
  arsaw07/future.ppt
• EmuLab site
• http//www.emulab.net/
• FlexLab A Realistic, Controlled , and Friendly
  Environment
• A Blueprint for Introducing Disruptive Technology
  into the Internet
• PlanetLab An Overlay Testbed for Broad Coverage
  Services
• Challenges and Opportunities in Internet Data
  Mining
• Reverse Engineering the Internet
• An Integrated Experimental Environment for
  Distributed Systems and Networks