University of Illinois at Chicago tronic Visualization Laboratory (EVL) Global Scale Tele-Immersion Network Performance Activities Jason Leigh, Oliver Yu, Linda Winkler, Alan Verlo, Tom DeFanti Yong-joo Cho, Ray Fang, Javier Girado, Liujia Hu, Tomoko Imai, Naveen Krishnaprasad, Michael Lewis, Ya Ju Lin, Dave Pape, Kyoung Park, Chris Scharver, Brenda Silva, Liang Wang Josh Eliason, Jinghua Ge, Eric He, Atul Nayak, Shalini Venkatamaran
31
Embed
University of Illinois at Chicago Electronic Visualization Laboratory (EVL) Global Scale Tele-Immersion Network Performance Activities Jason Leigh, Oliver.
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
University of Illinois at ChicagoElectronic Visualization Laboratory (EVL)
Global Scale Tele-Immersion Network Performance Activities
Jason Leigh, Oliver Yu, Linda Winkler, Alan Verlo, Tom DeFanti
Yong-joo Cho, Ray Fang, Javier Girado, Liujia Hu, Tomoko Imai, Naveen Krishnaprasad, Michael Lewis, Ya Ju Lin, Dave Pape, Kyoung Park, Chris Scharver,
Brenda Silva, Liang WangJosh Eliason, Jinghua Ge, Eric He, Atul Nayak, Shalini Venkatamaran
University of Illinois at ChicagoElectronic Visualization Laboratory (EVL)
Common Characteristics of Teleimmersive Applications
CAVE
Im m ersaDesk
Tele-Im m ersionClients
Tele-Im m ersionServer
Rem ote Data &Com putation
Services
Compute or DatabaseQuery Spawned by
Tele-ImmersionClient and M anaged
by Tele-ImmersionServer
University of Illinois at ChicagoElectronic Visualization Laboratory (EVL)
Characterization of Tele-Immersive Streams
Estimated bandwidth
(bits/s)
DiffServ
Types BurstinessLatency sensitive
Jitter sensitive
Error sensitive
UDP avatar 6K x n
(15fps)
Interactive Real-time
Constant Y Y N
UDP audio stream
64K x n Brief Y Y N
UDP video stream
10M
(2-way only)Constant Y Y YN
UDP stream
With Playback dependsNon-
interactive Real-time
Constant Y N YN
TCP control data 7K x n Reliable Brief YN YN Y
TCP bulk datadepends
Best Effort or Deadline Delivery
Sustained burst
N N Y
University of Illinois at ChicagoElectronic Visualization Laboratory (EVL)
Network Research
University of Illinois at ChicagoElectronic Visualization Laboratory (EVL)
Maximizing Bandwidth Utilization over Long Fat Networks
• Even if QoS via DiffServ or IntServ is available, it still does not solve the Long Fat Network problem
• Problem is small TCP window sizes (well known problem but still no widely accepted solution)
• On SGI’s change in window size requires kernel rebuild
• Size of window should be set to current available BW of the network
• CAVERNsoft’s Parallel Socket Striping works well but is considered “irresponsible” use of networks
University of Illinois at ChicagoElectronic Visualization Laboratory (EVL)
64K Window SizeAmsterdam to Chicago
Achieved bandwidth while transmitting 50M from Amsterdam to Chicago over a 45Mbps link
0
10
20
30
40
1 3 5 7 9 11
13
15
17
19
21
23
25
27
29
Number of Sockets
Ba
nd
wid
th (
Mb
ps)
Bursty as max bw reachedbut performance is still good
University of Illinois at ChicagoElectronic Visualization Laboratory (EVL)
64K Window SizeCERN to EVL
Plot of Average Achievable Bandwidth vs # of Parallel TCP Sockets Used to Deliver a 50M File from Switzerland (CERN) to Chicago over a 45Mbps link
University of Illinois at ChicagoElectronic Visualization Laboratory (EVL)
Anomalies
• Theoretical BW from EVL to SARA is 100Mbps• Netperf UDP shows reasonable performance:
– EVL to SARA 85Mbps– SARA to EVL 65Mbps (5 more hops via Abilene)
• Netperf and Parallel sockets TCP shows only:– 30Mbps
• Perhaps due to asymmetric tcp window size settings?
• Argument for UDP-based schemes?E.g. Forward Error Correction
University of Illinois at ChicagoElectronic Visualization Laboratory (EVL)
Forward Error Correction scheme for low-latency delivery of error sensitive data
• Transmit error correction data over high bandwidth networks that can be used for correcting UDP streams to achieve lower latency than TCP but higher reliability.
• Transmit error correction data to improve quality of streamed video by correcting for lost packets.
• Not intended for bulk data transfer but in light of TCP results this might hold some promise.
University of Illinois at ChicagoElectronic Visualization Laboratory (EVL)
FEC Experiments
• EVL to SARA- Amsterdam (45Mb/s 100ms RT latency)
• Broader Ques:– Can FEC provide a benefit? How much?– Tradeoff between redundancy and benefit?
• Specific Ques:– TCP vs UDP vs FEC/UDP– How much jitter does FEC introduce?– High thru put UDP vs FEC/UDP to observe loss &
recovery
University of Illinois at ChicagoElectronic Visualization Laboratory (EVL)
`Latency of transmitting 100 packets underUDP, TCP, FEC/UDP with 3:1 redundancy.
0
50
100
150
200
250
300
350
400
0 500 1000 1500 2000 2500
Packet size in bytes
1-w
ay la
ten
cy in
ms
UDP
TCP
FEC over UDP
FEC greatest benefit is in small packets.
Larger packets impose greater overhead.
As redundancy decreases FEC approaches UDP.
goal
University of Illinois at ChicagoElectronic Visualization Laboratory (EVL)
Jitter for UDP, TCP and FEC over UDPMoving average (over 20 successive data points) of deviations of Short Term Latency (also over 20
successive data points)
0
2
4
6
8
10
12
14
1 5 9
13
17
21
25
29
33
37
41
45
49
53
57
61
65
69
73
77
Jit
ter
UDP
TCP
FEC/UDP
G o a l
University of Illinois at ChicagoElectronic Visualization Laboratory (EVL)
Packet Loss over UDP vs FEC/UDP between Chicago & Amsterdam
Data Rate(bits/s)
Packet Size(Bytes)
Packet Loss Rate in UDP (%)
Packet Loss Rate in FEC over UDP (%)
1M 128 0.4 0
1M 256 0.2 0
1M 1024 0.2 0
10M 128 30 4
10M 256 25 3
10M 1024 21 1.5
UDP
UDP
FEC
University of Illinois at ChicagoElectronic Visualization Laboratory (EVL)
Human Factors in Tele-Immersion
University of Illinois at ChicagoElectronic Visualization Laboratory (EVL)
Collaborative Coordination Experiments between Chicago and Singapore
CAVE to CAVE (STAR TAP)• Audio via Phone call• Scramnet (adjustable latency, 0 jitter)• LAN Ethernet (~ 10ms)• Local ISDN (~ 200ms)• STAR TAP (~ 250ms)• Predict STAR TAP similar to performance over ISDN
University of Illinois at ChicagoElectronic Visualization Laboratory (EVL)
Collaborative Coordination Experiments between Chicago and Singapore
0
1
2
3
4
5
6
7
8
9
Scramnet 10ms LAN Ethernet ~10ms Scramnet 200ms ISDN ~200ms
Completion Time (s)
Collisions
• 200ms RTT is the threshold where performance begins to suffer• Roughly RTT to Asia. Results to Singapore similar to local ISDN
200ms RTT with 0 jitter is same as 10ms RTT with 7ms jitter
University of Illinois at ChicagoElectronic Visualization Laboratory (EVL)
DiffServ Bandwidth
0
5
10
15
20
25
30
1 15
29
43
57
71
85
99
113
127
141
155
169
183
197
211
225
239
Time (s)
Ba
nd
wid
th (
Mb
ps)
DiffServ Latency
0
10
20
3040
50
60
70
1 15
29
43
57
71
85
99
113
127
141
155
169
183
197
211
225
239
Time (s)
1 w
ay L
ate
ncy (
ms)
DiffServ Packet Loss
0
100
200
300
400
500
1 15
29
43
57
71
85
99
113
127
141
155
169
183
197
211
225
239
Time (s)
Pa
cke
t L
oss (
pa
cke
ts/s
)
25Mbps80Mbps
ANL
EVL
42Mbps 42Mbps
100Mbps 100Mbps
100Mbps 100Mbps
fore back
Bandwidth recovery good
Latency recovery good
Small packet loss
DiffServ Experiment 1+ background + DiffServ
x
x
xx
University of Illinois at ChicagoElectronic Visualization Laboratory (EVL)
DiffServ Experiment 2DiffServ Bandwidth
0
5
10
15
20
25
1 11
21
31
41
51
61
71
81
91
101
111
121
131
141
151
161
Time (s)
Ba
nd
wid
th (
Mb
ps)
DiffServ Latency
0
50
100
150
200
250
300
1 11
21
31
41
51
61
71
81
91
101
111
121
131
141
151
161
Time (s)
1 w
ay L
ate
ncy (
ms)
DiffServ Packet Loss
0200400600800
1000120014001600
1 10
19
28
37
46
55
64
73
82
91
100
109
118
127
136
145
154
163
Time (s)
Pa
cke
t L
oss
(p
acke
ts/s
)
25Mbps80Mbps
ANL
EVL
42Mbps 42Mbps
100Mbps 100Mbps
100Mbps 100Mbps
fore
back
Bandwidth recovery good
Latency recovery not good
Packet loss double
+ background + DiffServ
x
x
x x
University of Illinois at ChicagoElectronic Visualization Laboratory (EVL)
Application of Research Results
CAVERNsoft G2
applications at iGrid 2000 in Yokohama
University of Illinois at ChicagoElectronic Visualization Laboratory (EVL)