Separating Network Striping Policy from Mechanism

1
Separating Network Striping Policy from Mechanism

• Asfandyar Qureshi
• and
• John Guttag

2
Overview

• Horde Networking Middleware
• Provides a simple and robust way for multi-stream
  applications to communicate over multiple
  channels with widely varying characteristics.
• Key problems addressed
• Allow applications to abstractly influence packet
  scheduling.
• Provide a mechanism that derives the appropriate
  packet TX schedules.

3
Talk Structure

• Motivation and Background
• Motivating Application
• Public Wireless Networks
• Network Striping
• WWAN Striping Challenges
• Packet Scheduling
• Horde Middleware
• Services Provided
• Objective Driven Scheduling
• Channel Management

4
Motivating Application

• Mobile Telemedicine (joint with MGH)
• Provide advanced remote diagnostics capabilities
• Allow doctors to examine patients in-transit
• What we want to send
• Unidirectional VIDEO (300 kbit/sec)
• Bidirectional AUDIO (64 kbit/sec)
• Low-rate Physiological data (EKG, Heart-rate,
  etc)

5
Mobile Telemedicine

• Communication requirements
• Sustained high throughput
• Low-latency
• Vehicular motion in an urban area
• Economics
• System must be viable to deploy and operate
• Approach
• Use COTS components and public carrier wireless
  communications infrastructure

6
Public Wireless Networks

• Wireless Wide-Area Data Networks are ubiquitous
  in urban Areas
• Multiple providers (T-mobile, Verizon, )
• Multiple technologies (GSM/GPRS, CDMA, )
• Providers have overlapping coverage
• Network QoS is not great and not stable
• Latencies are high and variable
• GPRS mean 560ms, stdev 100ms
• CDMA mean 320ms, stdev 80ms
• Upload bandwidths (moving) are low and variable
• GPRS mean 20 kbps, stdev 5
• CDMA mean 120 kbps, stdev 20

7
Public Wireless Networks

• Wireless Wide-Area Data Networks are ubiquitous
  in urban Areas
• Multiple providers (T-mobile, Verizon, )
• Multiple technologies (GSM/GPRS, CDMA, )
• Providers have overlapping coverage
• Network QoS is not great and not stable
• Latencies are high and variable
• GPRS mean 560ms, stdev 100ms
• CDMA mean 320ms, stdev 80ms
• Upload bandwidths (moving) are low and variable
• GPRS mean 20 kbps, stdev 5
• CDMA mean 120 kbps, stdev 20

8
Network Striping

• Stripe Application Data across Multiple Network
  Channels
• Take data units from application and send them in
  some order over the channels.

A
B
N channels
M streams
M streams
9
Challenges I Application

• Bandwidth Limited Application
• Can send more data than network can accommodate
• Different Types of Streams with Dissimilar Needs
• Video, Audio, Bulk Data
• Different Network QoS Sensitivities
• Want applications to be independent of the number
  and types of channels

10
Challenges II Networks

• Network Channels are Dissimilar
• CDMA has 6x the bandwidth of GPRS
• Different technologies, many idiosyncrasies
• Network QoS Varies in Time
• Packet latency stdevs are 80
• Number of Channels Varies
• Motion makes this problem worse
• Forward Compatibility
• Different wireless network technologies will
  eventually be deployed

11
Horde Design Goals

• A Wireless Striping Middleware
• Can be useful to expose aspects of the striping
  operation to applications
• Develop a Powerful Set of Abstractions
• Make it easy to build diverse applications
• Dont sacrifice performance
• Support a heterogeneous set of unstable wireless
  channels
• Modularity is important

12
Network Striping Scheduling
APPLICATION
Different Requirements (Bandwidth QoS)
STREAMS
Different types of Service (Bandwidth QoS)
INTERFACES
NETWORK SERVICES
13
Network Striping Scheduling
APPLICATION
Different Requirements (Bandwidth QoS)
STREAMS
Different types of Service (Bandwidth QoS)
INTERFACES
NETWORK SERVICES
14
Network Striping Scheduling
APPLICATION
Different Requirements (Bandwidth QoS)
STREAMS
HORDE
Middleware
Different types of Service (Bandwidth QoS)
INTERFACES
NETWORK SERVICES
15
Network Striping Scheduling
APPLICATION
Different Requirements (Bandwidth QoS)
DATA UNITS
HORDE
Middleware
Different types of Service (Bandwidth QoS)
TX SLOTS
NETWORK SERVICES
16
Packet Scheduling (simple)

• Randomized Round Robin
• Stripe four streams with the same throughputs
  across one CDMA and three GPRS channels
• All streams get the same QoS
• Should all streams get the same QoS?

17
Packet Scheduling (better)

• Objective Driven Scheduling
• Packet scheduler incorporates application
  specific information (Streams 2 and 4 are video
  streams)
• Optimizes the division of the shared network
  resource based on stream sensitivities

18
Horde Middleware Overview
APPLICATION
HORDE
API
Packet Scheduler
Network Channel Management
O/S NETWORK SERVICES
19
Horde Middleware Overview
APPLICATION
HORDE
API
Packet Scheduler
Network Channel Management
O/S NETWORK SERVICES
20
Horde Middleware

• Provides a Number of Services
• Schedule data streams over channels
• Applications can modulate per stream QoS
• Applications abstractly influence striping policy
• Network channel congestion control
• Stream flow control
• Initial Implementation
• User-space
• Event driven API
• Similar to Congestion Manager OSDI 00

21
Horde Middleware

• Provides a Number of Services
• Schedule data streams over channels
• Applications can modulate per stream QoS
• Applications abstractly influence striping policy
• Network channel congestion control
• Stream flow control
• Initial Implementation
• User-space
• Event driven API
• Similar to Congestion Manager OSDI 00

22
QoS Modulation

• Streams have varying QoS sensitivities
• QoS is multidimensional
• Bandwidth
• Latency
• Loss and loss correlation
• Want to allow applications to express stream QoS
  sensitivities
• Interface must be flexible
• Applications must be easy to program

23
Application Utility

• UTILITY When an application sends data, it
  receives some utility from the consumption of its
  data at another host
• Total value derived from network service, minus
  cost
• Utility can be affected by the type of the data
  unit (e.g., a video I-frame) or the network-QoS
  for the data unit
• Similar to Microeconomics net consumer utility
• Utility Function
• We use a notion of an application-specified
  utility function
• This function allows the packet scheduler to
  abstractly determine application sensitivities
• Pick Schedules that Maximize Utility

24
Horde QoS Objectives

• Applications express QoS objectives
• Objectives define QoS constraints on streams
• Each objective defines a QoS goal and how the
  achievement of that goal adds to, or subtracts
  from, overall application utility
• E.g., an objective for a video stream could
  express that I-frame ADUs should have lower loss
  than P-frame ADUs

25
Horde QoS Objectives

• Objectives are
• Modular
• Correspond to QoS Dimensions
• Can refer to such things as expected latency and
  loss for an ADU or stream
• Independent of the number of channels
• The number and the nature of active network
  channels is likely to vary in a mobile
  application
• Expressed using a specification language

26
Expressing Objectives

• Stream audio1 values an average latency less than
  one second

objective context streamfoo
stream_id audio1 goal
foolatency_ave lt 1000 utility foo 100

27
Expressing Objectives

• Stream audio1 values an average latency less than
  one second

objective context streamfoo
stream_id audio1 goal
foolatency_ave lt 1000 utility foo 100

28
Expressing Objectives

• Stream audio1 values an average latency less than
  one second

objective context streamfoo
stream_id audio1 goal
foolatency_ave lt 1000 utility foo 100

29
Expressing Objectives

• Stream audio1 values an average latency less than
  one second

objective context streamfoo
stream_id audio1 goal
foolatency_ave lt 1000 utility foo 100

30
Expressing Objectives

• I-frames should have lower loss than others

objective context adufoo
(stream_id video1)
(frame_type I) adubar
(stream_id video1)
(frame_type ! I) goal
prob(foolost?) lt prob(barlost?) utility
foo 100
31
Objective Driven Scheduling

• Find high expected utility TX schedules
• Interpret set of expressed objectives
• Search over sub-space of possible TX schedules
• For example a random walk

32
Horde Middleware Overview
APPLICATION
HORDE
API
Packet Scheduler
Network Channel Management
O/S NETWORK SERVICES
33
Horde Middleware Overview
APPLICATION
HORDE
API
Packet Scheduler
Network Channel Management
O/S NETWORK SERVICES
34
Channel Management
Packet Scheduler
Network Channel Management
O/S NETWORK SERVICES

• Congestion Control
• Limits the schedulers sending rate
• Channel Bandwidth and QoS Estimation
• Prediction (near future)
• Needed to make good scheduling decisions

35
Channel Managers
Packet Scheduler
M1
M2
MN
O/S NETWORK SERVICES

• A single channel manager instance for each active
  network interface
• Different channel manager implementations for
  different network types
• Example one implementation to deal with CDMA,
  one for GPRS, and another one for 802.11

36
Transmission Slots

• For Scheduler, channels are sources of TxSlots
• Scheduler can abstract away channel specific
  idiosyncrasies
• TxSlots grant TX capabilities
• Scheduler collects slots and maps data to each
  slot
• Encapsulate expected QoS
• Have fields for expected latency, loss
  probability, etc

Scheduler
S
Mk
O/S
37
Phantom TX Slots

• Short-term channel QoS prediction boosts
  scheduler accuracy
• E.g., If a low-latency slot is likely to be
  available shortly, defer scheduling an urgent
  packet.
• Phantom TxSlots allow scheduler to factor in
  channel-specific predictions
• Phantoms are TxSlots that cant be used to send
  data
• Phantoms have confidence levels

Scheduler
S
Mk
O/S
38
Summary

• Goal was to build a flexible network striping
  middleware for WWANs
• Handle both channel and stream heterogeneity
• Two key abstractions
• Objectives
• Allow abstract manipulation of striping policy
• Transmission Slots
• Decouple scheduler from channel specific
  idiosyncrasies

39
Conclusions

• Using Horde it is possible to express rich
  objectives
• Rich enough for many interesting apps
• Maybe richer than needed
• Very simple schedulers can produce better
  schedules than would be produced in the absence
  of objectives
• Objective driven scheduling accounts for
  different stream QoS sensitivities