Presentation

Delay Analysis of Real-Time Data
Dissemination
Gidon Gershinsky, Avi Harpaz, Nir Naaman, Harel Paz,
Konstantin Shagin
IBM Haifa Labs
© 2007 IBM Corporation
IBM Labs in Haifa
Motivation
Performance guarantees and predictable task execution-time are
becoming increasingly important
Distributed real-time applications require a predictable form of
communication between the different components
Need to use a messaging system that delivers messages according to
their timeliness requirements
© 2007 IBM Corporation
IBM Labs in Haifa
The Messaging System Model
We consider a messaging transport that provides real-time quality of
service
One-to-many reliable message dissemination
The network transport is assumed to be unreliable
The messaging transport implements a reliable dissemination
protocol
Sender-initiated protocol (ACK based): upon receiving a packet, a
receiver sends an ACK to the sender
Receiver-initiated protocol (NAK based): a receiver sends a NAK for
missing packets
© 2007 IBM Corporation
IBM Labs in Haifa
This Work
We investigate messages delay: the delay between a message
submission (generation) at the sender and its delivery to all receivers
Provide an analytic model for congestion in data dissemination protocols
A rate control mechanism limits the sender’s transmission rate
Investigate transmission rate effect on message delivery latency
Main aspect examined: receiver side congestion
We consider the processing overhead of a discarded packet
Use a simulation to study additional effects
Evaluate the on-time delivery probability of a message
© 2007 IBM Corporation
IBM Labs in Haifa
Referred Applications
We investigate messages delay of 2 types of applications:
Bulk-data applications- very large messages
A message is converted into many packets
Short-message applications- bursty nature
An ON/OFF source
Alternates between ON and OFF periods
During ON periods packets are generated at a high rate
During OFF periods no packets are generated
© 2007 IBM Corporation
IBM Labs in Haifa
Results
Computational & simulation results
Processing times & network parameters are based on experiments
Analysis uses expectations
Simulation employs uniform distributions
Time unit = the time is takes a receiver to handle a single packet
The receiver rate = 1.0 packets per time unit
Examine various transmission rates
Low transmission rate ( 1.0): no congestion
High transmission rate ( > 1.0): congestion occurs
Discarding overhead is set to 0.3 time units
Network loss = 0.01
© 2007 IBM Corporation
IBM Labs in Haifa
Computational Results: Bulk-Data
# of Packet Transmissions as Function of Transmission Rate (ACK-based)
90
1 receiver
80
10 receivers
100 receivers
70
1000 receivers
60
50
40
30
20
10
3.
5
3.
3
3.
1
2.
9
2.
7
2.
5
2.
3
2.
1
1.
9
1.
7
1.
5
1.
3
1.
1
0.
9
0.
7
0.
5
0.
3
0
0.
1
number of packet transmissions
100
normalized transmission rate
© 2007 IBM Corporation
IBM Labs in Haifa
Computational Results: Bulk-Data
Delivery Time as Function of Transmission Rate (ACK-based)
3.50E+05
1 receiver
3.00E+05
100 receivers
10 receivers
1000 receivers
2.50E+05
2.00E+05
1.50E+05
1.00E+05
5.00E+04
1
3
5
7
9
1
3
5
7
9
1
3
5
7
9
1
3
5
0.
0.
0.
0.
1.
1.
1.
1.
1.
2.
2.
2.
2.
2.
3.
3.
3.
0.00E+00
0.
delivery time (in time units)
4.00E+05
normalize d transmission rate
© 2007 IBM Corporation
IBM Labs in Haifa
Computational Results: Bulk-Data
Delivery Time as Function of Transmission Rate (ACK-based)
4.00E+05
3.50E+05
1 receiver
3.00E+05
100 receivers
d elive ry tim e (in tim e u n its)
10 receivers
1000 receivers
2.50E+05
1 receiver
3.50E+05
2.00E+05
1.50E+05
1.00E+05
10 receivers
100 receivers
3.00E+05
1000 receivers
2.50E+05
2.00E+05
1.50E+05
1.00E+05
5.00E+04
5.00E+04
9
1
3
5
7
9
1
3
5
1.
2.
2.
2.
2.
2.
3.
3.
3.
5
3.
7
3
3.
5
1
3.
1.
9
2.
3
7
2.
1.
5
2.
1.
3
2.
1
1
2.
9
9
1.
1.
7
1.
7
5
1.
0.
3
1.
0.
1
1.
5
9
0.
3
7
0.
0.
5
1
3
0.
0.
1
0.
normalize d transmission rate
0.
0.00E+00
0.00E+00
0.
d e liv e ry tim e (in tim e u n its )
4.00E+05
normalized transmission rate
Discard overhead is ignored
© 2007 IBM Corporation
IBM Labs in Haifa
Computational Results: Bulk-Data
Delivery Time as Function of Transmission Rate (NAK-based)
9.00E+05
1 receiver
8.00E+05
10 receivers
100 receivers
7.00E+05
1000 receivers
6.00E+05
5.00E+05
4.00E+05
3.00E+05
2.00E+05
1.00E+05
1
3
5
7
9
1
3
5
7
9
1
3
5
7
9
1
0.
0.
0.
0.
1.
1.
1.
1.
1.
2.
2.
2.
2.
2.
3.
0.00E+00
0.
delivery time (in time units)
1.00E+06
normalized transmission rate
© 2007 IBM Corporation
IBM Labs in Haifa
Results- Short Messages
Duration
(time units)
Message
generation rate
(per time unit)
ON period
50
3
OFF period
250
0
© 2007 IBM Corporation
IBM Labs in Haifa
Simulation: ON/OFF Model
Latency Measurements (10 receivers)
600
rate .5
latency (in time units)
500
400
phase 2
300
phase 1
200
100
0
1
73
145 217 289 361 433 505 577 649 721 793 865 937 1009 1081 1153 1225 1297 1369 1441
packet
number
© 2007 IBM Corporation
IBM Labs in Haifa
Simulation: ON/OFF Model
Delivery by Deadline Probability (Different Receiver Sets)
1.2
1 receiver
1
10 receivers
1 receiver
100 receivers
0.6
0.4
10 receivers
100 receivers
0.2
7
9
1
3
5
7
9
1
3
1.
2.
2.
2.
2.
2.
3.
3.
1
1.
1.
9
0.
5
7
0.
1.
5
0.
3
3
0.
1.
1
0
0.
probability
0.8
normalized transmission rate
© 2007 IBM Corporation
IBM Labs in Haifa
Summary
We study aspects related to real-time message dissemination
Applications that generate large messages and short messages
Sender-initiated and receiver-initiated protocols
Congestion at the receivers
The analytic model draws the dependency between transmission
rate limit and the expected message latency
The simulation indicates that the analysis’ simplifications have a
negligible effect
The analysis can be incorporated in a transmission rate control logic to
attain high probability of timely delivery and support stricter real-time
requirements
© 2007 IBM Corporation