research on cloud gaming: status and perspectives
TRANSCRIPT
About gaming
59% of Americans play video games39% of gamers are 36+ years old48% of gamers are female
The global revenue will hit $100 billion by 2018More than Hollywood movie industryGaming is 20% of iPhone downloaded apps
147.000 jobs are related to gaming in 2014 in US
2 / 36 Gwendal Simon Cloud Gaming
About gaming
59% of Americans play video games39% of gamers are 36+ years old48% of gamers are female
The global revenue will hit $100 billion by 2018More than Hollywood movie industryGaming is 20% of iPhone downloaded apps
147.000 jobs are related to gaming in 2014 in US
2 / 36 Gwendal Simon Cloud Gaming
About gaming
59% of Americans play video games39% of gamers are 36+ years old48% of gamers are female
The global revenue will hit $100 billion by 2018More than Hollywood movie industryGaming is 20% of iPhone downloaded apps
147.000 jobs are related to gaming in 2014 in US
2 / 36 Gwendal Simon Cloud Gaming
About gaming
59% of Americans play video games39% of gamers are 36+ years old48% of gamers are female
The global revenue will hit $100 billion by 2018More than Hollywood movie industryGaming is 20% of iPhone downloaded apps
147.000 jobs are related to gaming in 2014 in US
2 / 36 Gwendal Simon Cloud Gaming
Gaming modular architecture
Gamer CommandInterface
Game Engine
Graphical Output
GameServer
Console/PC market based on power-ful hardware. Target hardcore gamers
PC + software Smartphone market based on appstore. Target casual gamers
smartphone + app
Cloud gaming based on the cloud andlightweight devices. Target everybody
cloud
home
3 / 36 Gwendal Simon Cloud Gaming
Gaming modular architecture
Gamer CommandInterface
Game Engine
Graphical Output
GameServer
Console/PC market based on power-ful hardware. Target hardcore gamers
PC + software
Smartphone market based on appstore. Target casual gamers
smartphone + app
Cloud gaming based on the cloud andlightweight devices. Target everybody
cloud
home
3 / 36 Gwendal Simon Cloud Gaming
Gaming modular architecture
Gamer CommandInterface
Game Engine
Graphical Output
GameServer
Console/PC market based on power-ful hardware. Target hardcore gamers
PC + software
Smartphone market based on appstore. Target casual gamers
smartphone + app
Cloud gaming based on the cloud andlightweight devices. Target everybody
cloud
home
3 / 36 Gwendal Simon Cloud Gaming
Gaming modular architecture
Gamer CommandInterface
Game Engine
Graphical Output
GameServer
Console/PC market based on power-ful hardware. Target hardcore gamers
PC + software
Smartphone market based on appstore. Target casual gamers
smartphone + app
Cloud gaming based on the cloud andlightweight devices. Target everybody
cloud
home
3 / 36 Gwendal Simon Cloud Gaming
Cloud gaming architecture in a nutshell
Data-CenterInternet
user actions
video stream
4 / 36 Gwendal Simon Cloud Gaming
Cloud gaming architecture in a nutshell
Data-CenterInternet
user actions
video stream
4 / 36 Gwendal Simon Cloud Gaming
Cloud gaming architecture in a nutshell
Data-CenterInternet
user actions
video stream
4 / 36 Gwendal Simon Cloud Gaming
Why does it make sense ?
For game developers :Better piracy control
Support more client platformsEasier to upgradeNew business offers (e.g. subscription-based)
For gamers :No more hardwarePlay games on thin clientsFollow-me gaming
5 / 36 Gwendal Simon Cloud Gaming
Why does it make sense ?
For game developers :Better piracy controlSupport more client platforms
Easier to upgradeNew business offers (e.g. subscription-based)
For gamers :No more hardwarePlay games on thin clientsFollow-me gaming
5 / 36 Gwendal Simon Cloud Gaming
Why does it make sense ?
For game developers :Better piracy controlSupport more client platformsEasier to upgrade
New business offers (e.g. subscription-based)
For gamers :No more hardwarePlay games on thin clientsFollow-me gaming
5 / 36 Gwendal Simon Cloud Gaming
Why does it make sense ?
For game developers :Better piracy controlSupport more client platformsEasier to upgradeNew business offers (e.g. subscription-based)
For gamers :No more hardwarePlay games on thin clientsFollow-me gaming
5 / 36 Gwendal Simon Cloud Gaming
Why does it make sense ?
For game developers :Better piracy controlSupport more client platformsEasier to upgradeNew business offers (e.g. subscription-based)
For gamers :No more hardware
Play games on thin clientsFollow-me gaming
5 / 36 Gwendal Simon Cloud Gaming
Why does it make sense ?
For game developers :Better piracy controlSupport more client platformsEasier to upgradeNew business offers (e.g. subscription-based)
For gamers :No more hardwarePlay games on thin clients
Follow-me gaming
5 / 36 Gwendal Simon Cloud Gaming
Why does it make sense ?
For game developers :Better piracy controlSupport more client platformsEasier to upgradeNew business offers (e.g. subscription-based)
For gamers :No more hardwarePlay games on thin clientsFollow-me gaming
5 / 36 Gwendal Simon Cloud Gaming
Business status
The pioneers have shut down :OnLive bankrupted in 2013. Too earlyStreamMyGame never took off.
Last game consoles integrate cloud gaming solutions :Sony acquired Gaikai (PS3 backward compatibility)Microsoft launched XBox Live Cloud a year later
New actors enter the market :NVidia released Nvidia Grid and Shield
6 / 36 Gwendal Simon Cloud Gaming
Business status
The pioneers have shut down :OnLive bankrupted in 2013. Too earlyStreamMyGame never took off.
Last game consoles integrate cloud gaming solutions :Sony acquired Gaikai (PS3 backward compatibility)Microsoft launched XBox Live Cloud a year later
New actors enter the market :NVidia released Nvidia Grid and Shield
6 / 36 Gwendal Simon Cloud Gaming
Business status
The pioneers have shut down :OnLive bankrupted in 2013. Too earlyStreamMyGame never took off.
Last game consoles integrate cloud gaming solutions :Sony acquired Gaikai (PS3 backward compatibility)Microsoft launched XBox Live Cloud a year later
New actors enter the market :NVidia released Nvidia Grid and Shield
6 / 36 Gwendal Simon Cloud Gaming
The main reasons for latency
Response time is the delay from anaction done at the user and the result ofthis action on the display
gatewayt1 Internett2 DC routert3 user inputprocessing
t4
game stateupdate
t5
graphicsrendering
t6
videoencoding
t7
networkdelivery
t1 + t2 + t3videodecoding
t8
latency due to network
latency due to game engine
latency due to video encoding and decoding
8 / 36 Gwendal Simon Cloud Gaming
The main reasons for latency
Response time is the delay from anaction done at the user and the result ofthis action on the display
gatewayt1
Internett2 DC routert3 user inputprocessing
t4
game stateupdate
t5
graphicsrendering
t6
videoencoding
t7
networkdelivery
t1 + t2 + t3videodecoding
t8
latency due to network
latency due to game engine
latency due to video encoding and decoding
8 / 36 Gwendal Simon Cloud Gaming
The main reasons for latency
Response time is the delay from anaction done at the user and the result ofthis action on the display
gatewayt1 Internett2
DC routert3 user inputprocessing
t4
game stateupdate
t5
graphicsrendering
t6
videoencoding
t7
networkdelivery
t1 + t2 + t3videodecoding
t8
latency due to network
latency due to game engine
latency due to video encoding and decoding
8 / 36 Gwendal Simon Cloud Gaming
The main reasons for latency
Response time is the delay from anaction done at the user and the result ofthis action on the display
gatewayt1 Internett2 DC routert3
user inputprocessing
t4
game stateupdate
t5
graphicsrendering
t6
videoencoding
t7
networkdelivery
t1 + t2 + t3videodecoding
t8
latency due to network
latency due to game engine
latency due to video encoding and decoding
8 / 36 Gwendal Simon Cloud Gaming
The main reasons for latency
Response time is the delay from anaction done at the user and the result ofthis action on the display
gatewayt1 Internett2 DC routert3 user inputprocessing
t4
game stateupdate
t5
graphicsrendering
t6
videoencoding
t7
networkdelivery
t1 + t2 + t3videodecoding
t8
latency due to network
latency due to game engine
latency due to video encoding and decoding
8 / 36 Gwendal Simon Cloud Gaming
The main reasons for latency
Response time is the delay from anaction done at the user and the result ofthis action on the display
gatewayt1 Internett2 DC routert3 user inputprocessing
t4
game stateupdate
t5
graphicsrendering
t6
videoencoding
t7
networkdelivery
t1 + t2 + t3videodecoding
t8
latency due to network
latency due to game engine
latency due to video encoding and decoding
8 / 36 Gwendal Simon Cloud Gaming
The main reasons for latency
Response time is the delay from anaction done at the user and the result ofthis action on the display
gatewayt1 Internett2 DC routert3 user inputprocessing
t4
game stateupdate
t5
graphicsrendering
t6
videoencoding
t7
networkdelivery
t1 + t2 + t3videodecoding
t8
latency due to network
latency due to game engine
latency due to video encoding and decoding
8 / 36 Gwendal Simon Cloud Gaming
The main reasons for latency
Response time is the delay from anaction done at the user and the result ofthis action on the display
gatewayt1 Internett2 DC routert3 user inputprocessing
t4
game stateupdate
t5
graphicsrendering
t6
videoencoding
t7
networkdelivery
t1 + t2 + t3videodecoding
t8
latency due to network
latency due to game engine
latency due to video encoding and decoding
8 / 36 Gwendal Simon Cloud Gaming
The main reasons for latency
Response time is the delay from anaction done at the user and the result ofthis action on the display
gatewayt1 Internett2 DC routert3 user inputprocessing
t4
game stateupdate
t5
graphicsrendering
t6
videoencoding
t7
networkdelivery
t1 + t2 + t3
videodecoding
t8
latency due to network
latency due to game engine
latency due to video encoding and decoding
8 / 36 Gwendal Simon Cloud Gaming
The main reasons for latency
Response time is the delay from anaction done at the user and the result ofthis action on the display
gatewayt1 Internett2 DC routert3 user inputprocessing
t4
game stateupdate
t5
graphicsrendering
t6
videoencoding
t7
networkdelivery
t1 + t2 + t3videodecoding
t8
latency due to network
latency due to game engine
latency due to video encoding and decoding
8 / 36 Gwendal Simon Cloud Gaming
The main reasons for latency
Response time is the delay from anaction done at the user and the result ofthis action on the display
gatewayt1 Internett2 DC routert3 user inputprocessing
t4
game stateupdate
t5
graphicsrendering
t6
videoencoding
t7
networkdelivery
t1 + t2 + t3videodecoding
t8
latency due to network
latency due to game engine
latency due to video encoding and decoding
8 / 36 Gwendal Simon Cloud Gaming
The main reasons for latency
Response time is the delay from anaction done at the user and the result ofthis action on the display
gatewayt1 Internett2 DC routert3 user inputprocessing
t4
game stateupdate
t5
graphicsrendering
t6
videoencoding
t7
networkdelivery
t1 + t2 + t3videodecoding
t8
latency due to network
latency due to game engine
latency due to video encoding and decoding
8 / 36 Gwendal Simon Cloud Gaming
The main reasons for latency
Response time is the delay from anaction done at the user and the result ofthis action on the display
gatewayt1 Internett2 DC routert3 user inputprocessing
t4
game stateupdate
t5
graphicsrendering
t6
videoencoding
t7
networkdelivery
t1 + t2 + t3videodecoding
t8
latency due to network
latency due to game engine
latency due to video encoding and decoding
8 / 36 Gwendal Simon Cloud Gaming
Game genregame genre
avatar-based games
first-person games third-person games
omnipresent games
10 / 36 Gwendal Simon Cloud Gaming
Game genregame genre
avatar-based games
first-person games third-person games
omnipresent games
10 / 36 Gwendal Simon Cloud Gaming
Game genregame genre
avatar-based games
first-person games third-person games
omnipresent games
10 / 36 Gwendal Simon Cloud Gaming
Game genregame genre
avatar-based games
first-person games third-person games
omnipresent games
10 / 36 Gwendal Simon Cloud Gaming
Game genregame genre
avatar-based games
first-person games third-person games
omnipresent games
10 / 36 Gwendal Simon Cloud Gaming
Game genregame genre
avatar-based games
first-person games third-person games
omnipresent games
10 / 36 Gwendal Simon Cloud Gaming
Game genregame genre
avatar-based games
first-person games third-person games
omnipresent games
10 / 36 Gwendal Simon Cloud Gaming
Game genregame genre
avatar-based games
first-person games third-person games
omnipresent games
10 / 36 Gwendal Simon Cloud Gaming
Genre, latency, and performance
The Claypool law of latency acceptance :genre first-person third-person omnipresent
lag in ms 100 500 1000
Various other factors to consider :Limitations in the classification (e.g. StarCraft)User control games (more or less immersive)Gamer skill level (from casual to professional)
11 / 36 Gwendal Simon Cloud Gaming
Genre, latency, and performance
The Claypool law of latency acceptance :genre first-person third-person omnipresent
lag in ms 100 500 1000
Various other factors to consider :Limitations in the classification (e.g. StarCraft)User control games (more or less immersive)Gamer skill level (from casual to professional)
11 / 36 Gwendal Simon Cloud Gaming
Game action taxonomy
. . .add the impact of the action
Toward priority classes for actions ?
12 / 36 Gwendal Simon Cloud Gaming
Game action taxonomy
. . .add the impact of the action
Toward priority classes for actions ?
12 / 36 Gwendal Simon Cloud Gaming
Latency target
The shorter the better
The best range is from 100 to 150 ms
Up to 300 ms is enough for some games
Action prioritization may be standardized
13 / 36 Gwendal Simon Cloud Gaming
Latency target
The shorter the better
The best range is from 100 to 150 ms
Up to 300 ms is enough for some games
Action prioritization may be standardized
13 / 36 Gwendal Simon Cloud Gaming
Latency target
The shorter the better
The best range is from 100 to 150 ms
Up to 300 ms is enough for some games
Action prioritization may be standardized
13 / 36 Gwendal Simon Cloud Gaming
Latency target
The shorter the better
The best range is from 100 to 150 ms
Up to 300 ms is enough for some games
Action prioritization may be standardized
13 / 36 Gwendal Simon Cloud Gaming
Video
Gamers have high expectations :High-resolution (1080p possible since PS3)High frame rate (60 fps is common)Complex, dynamic scenes
Screencasting technologies are booming :Capture, encode, and packetizee.g., Open Broadcast Software
15 / 36 Gwendal Simon Cloud Gaming
Video
Gamers have high expectations :High-resolution (1080p possible since PS3)High frame rate (60 fps is common)Complex, dynamic scenes
Screencasting technologies are booming :Capture, encode, and packetizee.g., Open Broadcast Software
15 / 36 Gwendal Simon Cloud Gaming
Video decoding/encoding
In the Rhizome platform :encoding capture im. convert im. encode frame stream frame total720p 6.4ms 4.4ms 2.8ms 0.2ms 13.8ms1080p 11.0ms 4.9ms 3.0ms 0.3ms 19.2ms
Use info from the game engine to encode videoReduce Motion Estimation (ME) using
Game object informationDepth map
Reduce encoding complexity usingObject priority (adaptive object encoding)
17 / 36 Gwendal Simon Cloud Gaming
Video decoding/encoding
In the Rhizome platform :encoding capture im. convert im. encode frame stream frame total720p 6.4ms 4.4ms 2.8ms 0.2ms 13.8ms1080p 11.0ms 4.9ms 3.0ms 0.3ms 19.2ms
Use info from the game engine to encode videoReduce Motion Estimation (ME) using
Game object informationDepth map
Reduce encoding complexity usingObject priority (adaptive object encoding)
17 / 36 Gwendal Simon Cloud Gaming
Video decoding/encoding
In the Rhizome platform :encoding capture im. convert im. encode frame stream frame total720p 6.4ms 4.4ms 2.8ms 0.2ms 13.8ms1080p 11.0ms 4.9ms 3.0ms 0.3ms 19.2ms
Use info from the game engine to encode videoReduce Motion Estimation (ME) using
Game object informationDepth map
Reduce encoding complexity usingObject priority (adaptive object encoding)
17 / 36 Gwendal Simon Cloud Gaming
Virtualization on GPU
Cloud management is based on Virtual MachinesMature technologies to create, shut down, migrateHardware consolidation : several VMs per machine
Poor performances on traditional GPUDegrades memory transfer (GPU bottleneck)Prevents Single-Instruction-Multiple-Data
Pass-through GPU enables better performances :One-to-one mapping between device and VMNew technologies coming (e.g., NVidia Grid)
19 / 36 Gwendal Simon Cloud Gaming
Virtualization on GPU
Cloud management is based on Virtual MachinesMature technologies to create, shut down, migrateHardware consolidation : several VMs per machine
Poor performances on traditional GPUDegrades memory transfer (GPU bottleneck)Prevents Single-Instruction-Multiple-Data
Pass-through GPU enables better performances :One-to-one mapping between device and VMNew technologies coming (e.g., NVidia Grid)
19 / 36 Gwendal Simon Cloud Gaming
Virtualization on GPU
Cloud management is based on Virtual MachinesMature technologies to create, shut down, migrateHardware consolidation : several VMs per machine
Poor performances on traditional GPUDegrades memory transfer (GPU bottleneck)Prevents Single-Instruction-Multiple-Data
Pass-through GPU enables better performances :One-to-one mapping between device and VMNew technologies coming (e.g., NVidia Grid)
19 / 36 Gwendal Simon Cloud Gaming
Virtualization on GPU : performances
Bare-Metal Xen 4.0 Bare-Metal Xen 4.10
100
200
300
2011 2014126.2
39.8
274.7 266.9
frames
pers
econ
d
Doom3
20 / 36 Gwendal Simon Cloud Gaming
Virtualization on GPU : performances
Bare-Metal Xen 4.0 Bare-Metal Xen 4.10
100
200
300
2011 2014126.2
39.8
274.7 266.9
84.5
51
153.7 153.4
frames
pers
econ
d
Doom3Unigine Sanctuary
20 / 36 Gwendal Simon Cloud Gaming
Game virtualization : consolidationnb. of instances Bare-metal (in fps) VM (in fps)
1 30.2 29.22 30.0 28.93 29.8 28.9
1 2 30
200
400
600
800
1,000
287
475
662
296
485
704
Nb. of game instances
Wat
ts
Bare-metalVirtual machines
21 / 36 Gwendal Simon Cloud Gaming
Game virtualization : consolidationnb. of instances Bare-metal (in fps) VM (in fps)
1 30.2 29.22 30.0 28.93 29.8 28.9
1 2 30
200
400
600
800
1,000
287
475
662
296
485
704
Nb. of game instances
Wat
tsBare-metalVirtual machines
21 / 36 Gwendal Simon Cloud Gaming
Latency : summary
Incompressible latency (without network) :
Video encoding and decoding : 40 msGame engine : from 15 to 50 ms
A lot of works for key latency gains here and there :Adaptive streaming in both resolution and fpsPrioritize actionsProvide metadata and information on game objectsShared processes for multiple game instances
22 / 36 Gwendal Simon Cloud Gaming
Latency : summary
Incompressible latency (without network) :Video encoding and decoding : 40 ms
Game engine : from 15 to 50 ms
A lot of works for key latency gains here and there :Adaptive streaming in both resolution and fpsPrioritize actionsProvide metadata and information on game objectsShared processes for multiple game instances
22 / 36 Gwendal Simon Cloud Gaming
Latency : summary
Incompressible latency (without network) :Video encoding and decoding : 40 msGame engine : from 15 to 50 ms
A lot of works for key latency gains here and there :Adaptive streaming in both resolution and fpsPrioritize actionsProvide metadata and information on game objectsShared processes for multiple game instances
22 / 36 Gwendal Simon Cloud Gaming
Latency : summary
Incompressible latency (without network) :Video encoding and decoding : 40 msGame engine : from 15 to 50 ms
A lot of works for key latency gains here and there :
Adaptive streaming in both resolution and fpsPrioritize actionsProvide metadata and information on game objectsShared processes for multiple game instances
22 / 36 Gwendal Simon Cloud Gaming
Latency : summary
Incompressible latency (without network) :Video encoding and decoding : 40 msGame engine : from 15 to 50 ms
A lot of works for key latency gains here and there :Adaptive streaming in both resolution and fpsPrioritize actionsProvide metadata and information on game objectsShared processes for multiple game instances
22 / 36 Gwendal Simon Cloud Gaming
To contribute to the research effort
Some open platforms :GamingAnywhere, the most mature
Uniquitous based on Unity 3D
Rhizome with integration of NVidia Grid
23 / 36 Gwendal Simon Cloud Gaming
To contribute to the research effort
Some open platforms :GamingAnywhere, the most mature
Uniquitous based on Unity 3D
Rhizome with integration of NVidia Grid
23 / 36 Gwendal Simon Cloud Gaming
To contribute to the research effort
Some open platforms :GamingAnywhere, the most mature
Uniquitous based on Unity 3D
Rhizome with integration of NVidia Grid
23 / 36 Gwendal Simon Cloud Gaming
Our objectives
Can the current cloud support cloud gaming ?
Which percentage of users can play cloud games ?1. In excellent conditions (40 ms network latency)2. In good conditions (80 ms network latency)
25 / 36 Gwendal Simon Cloud Gaming
Our objectives
Can the current cloud support cloud gaming ?Which percentage of users can play cloud games ?
1. In excellent conditions (40 ms network latency)2. In good conditions (80 ms network latency)
25 / 36 Gwendal Simon Cloud Gaming
Our objectives
Can the current cloud support cloud gaming ?Which percentage of users can play cloud games ?1. In excellent conditions (40 ms network latency)
2. In good conditions (80 ms network latency)
25 / 36 Gwendal Simon Cloud Gaming
Our objectives
Can the current cloud support cloud gaming ?Which percentage of users can play cloud games ?1. In excellent conditions (40 ms network latency)2. In good conditions (80 ms network latency)
25 / 36 Gwendal Simon Cloud Gaming
Our population sample
A modified BitTorrent crawler in Dec. 2012 :Collected 2,504 IP addresses in USMost likely machines for home entertainment
26 / 36 Gwendal Simon Cloud Gaming
Our population sample
A modified BitTorrent crawler in Dec. 2012 :Collected 2,504 IP addresses in USMost likely machines for home entertainment
26 / 36 Gwendal Simon Cloud Gaming
Latency to EC2 : experiment
Amazon datacenters (Portland, Virginia, North Calif.)
For each client (IP address)1. Measure latency from each DC2. Pick the shortest latency3. Do it ten times
27 / 36 Gwendal Simon Cloud Gaming
Latency to EC2 : experiment
Amazon datacenters (Portland, Virginia, North Calif.)
For each client (IP address)1. Measure latency from each DC
2. Pick the shortest latency3. Do it ten times
27 / 36 Gwendal Simon Cloud Gaming
Latency to EC2 : experiment
Amazon datacenters (Portland, Virginia, North Calif.)
For each client (IP address)1. Measure latency from each DC2. Pick the shortest latency
3. Do it ten times
27 / 36 Gwendal Simon Cloud Gaming
Latency to EC2 : experiment
Amazon datacenters (Portland, Virginia, North Calif.)
For each client (IP address)1. Measure latency from each DC2. Pick the shortest latency3. Do it ten times
27 / 36 Gwendal Simon Cloud Gaming
Latency to EC2 : results
0 20 40 60 80 100 120 140 1600
0.20.40.60.81
median latency (in ms)
ratio
ofcoveredusers
28 / 36 Gwendal Simon Cloud Gaming
Latency to EC2 : results
0 20 40 60 80 100 120 140 1600
0.20.40.60.81
median latency (in ms)
ratio
ofcoveredusers
28 / 36 Gwendal Simon Cloud Gaming
Latency to EC2 : results
0 20 40 60 80 100 120 140 1600
0.20.40.60.81
two thirds of gamers ingood conditions
one third of gamers inexcellent conditions
one tenth of gamerscannot play at all
median latency (in ms)
ratio
ofcoveredusers
28 / 36 Gwendal Simon Cloud Gaming
What if more DCs are used
44 PlanetLabs nodes (all over the US)
For each client (IP address) :1. Measure latency from each PlanetLabs DC2. Do it ten times
For each k-subset of DCs :1. Choose for each client the shortest latency2. Sum all shortest latencies
Then pick the best k-subset of DCs
29 / 36 Gwendal Simon Cloud Gaming
What if more DCs are used
44 PlanetLabs nodes (all over the US)For each client (IP address) :1. Measure latency from each PlanetLabs DC
2. Do it ten times
For each k-subset of DCs :1. Choose for each client the shortest latency2. Sum all shortest latencies
Then pick the best k-subset of DCs
29 / 36 Gwendal Simon Cloud Gaming
What if more DCs are used
44 PlanetLabs nodes (all over the US)For each client (IP address) :1. Measure latency from each PlanetLabs DC2. Do it ten times
For each k-subset of DCs :1. Choose for each client the shortest latency2. Sum all shortest latencies
Then pick the best k-subset of DCs
29 / 36 Gwendal Simon Cloud Gaming
What if more DCs are used
44 PlanetLabs nodes (all over the US)For each client (IP address) :1. Measure latency from each PlanetLabs DC2. Do it ten times
For each k-subset of DCs :
1. Choose for each client the shortest latency2. Sum all shortest latencies
Then pick the best k-subset of DCs
29 / 36 Gwendal Simon Cloud Gaming
What if more DCs are used
44 PlanetLabs nodes (all over the US)For each client (IP address) :1. Measure latency from each PlanetLabs DC2. Do it ten times
For each k-subset of DCs :1. Choose for each client the shortest latency2. Sum all shortest latencies
Then pick the best k-subset of DCs
29 / 36 Gwendal Simon Cloud Gaming
What if more DCs are used
44 PlanetLabs nodes (all over the US)For each client (IP address) :1. Measure latency from each PlanetLabs DC2. Do it ten times
For each k-subset of DCs :1. Choose for each client the shortest latency2. Sum all shortest latencies
Then pick the best k-subset of DCs
29 / 36 Gwendal Simon Cloud Gaming
More DCs : results
5 10 15 20 250
0.2
0.4
0.6
0.8
1
EC2 80 ms
EC2 40 ms
number of datacenters
ratio
ofcoveredusers
30 / 36 Gwendal Simon Cloud Gaming
More DCs : results
5 10 15 20 250
0.2
0.4
0.6
0.8
1
EC2 80 ms
EC2 40 ms
number of datacenters
ratio
ofcoveredusers
80 ms latency threshold40 ms latency threshold
30 / 36 Gwendal Simon Cloud Gaming
Toward a hybrid DC-P2P infrastructure
Out of the 2,504 clients :1,000 are edge servers1,504 are clients
For each client :1. Measure latency to each Amazon DC2. Measure latency to all edge-servers3. Pick the shortest latency4. Do it ten times
31 / 36 Gwendal Simon Cloud Gaming
Toward a hybrid DC-P2P infrastructure
Out of the 2,504 clients :1,000 are edge servers1,504 are clients
For each client :1. Measure latency to each Amazon DC
2. Measure latency to all edge-servers3. Pick the shortest latency4. Do it ten times
31 / 36 Gwendal Simon Cloud Gaming
Toward a hybrid DC-P2P infrastructure
Out of the 2,504 clients :1,000 are edge servers1,504 are clients
For each client :1. Measure latency to each Amazon DC2. Measure latency to all edge-servers
3. Pick the shortest latency4. Do it ten times
31 / 36 Gwendal Simon Cloud Gaming
Toward a hybrid DC-P2P infrastructure
Out of the 2,504 clients :1,000 are edge servers1,504 are clients
For each client :1. Measure latency to each Amazon DC2. Measure latency to all edge-servers3. Pick the shortest latency
4. Do it ten times
31 / 36 Gwendal Simon Cloud Gaming
Toward a hybrid DC-P2P infrastructure
Out of the 2,504 clients :1,000 are edge servers1,504 are clients
For each client :1. Measure latency to each Amazon DC2. Measure latency to all edge-servers3. Pick the shortest latency4. Do it ten times
31 / 36 Gwendal Simon Cloud Gaming
Hybrid DC-P2P : results
0 20 40 60 80 100 120 140 1600
0.2
0.4
0.6
0.8
1
median latency (in ms)
ratio
ofcoveredusers
Amazon multi-DC infra.
32 / 36 Gwendal Simon Cloud Gaming
Hybrid DC-P2P : results
0 20 40 60 80 100 120 140 1600
0.2
0.4
0.6
0.8
1
median latency (in ms)
ratio
ofcoveredusers
Amazon multi-DC infra.Hybrid DC-P2P infra.
32 / 36 Gwendal Simon Cloud Gaming
Hybrid DC-P2P : results
+105%
+28%
+12%
0 20 40 60 80 100 120 140 1600
0.2
0.4
0.6
0.8
1
median latency (in ms)
ratio
ofcoveredusers
Amazon multi-DC infra.Hybrid DC-P2P infra.
32 / 36 Gwendal Simon Cloud Gaming
Network latency : summary
Not all users will be served
100 ms overall lag needs a lot of servers at the edges
Game placement on servers will be key
33 / 36 Gwendal Simon Cloud Gaming
Network latency : summary
Not all users will be served
100 ms overall lag needs a lot of servers at the edges
Game placement on servers will be key
33 / 36 Gwendal Simon Cloud Gaming
Network latency : summary
Not all users will be served
100 ms overall lag needs a lot of servers at the edges
Game placement on servers will be key
33 / 36 Gwendal Simon Cloud Gaming
Main researchers in the areaKuan-Ta Chen (Sinica - Taiwan)
Shervin Shirmohammadi (Ottawa - Canada)
Mark Claypool (Worcester - UK)
Jiangchuan Liu (Simon Fraser - Canada)
Carsten Griwodz (Simula - Norway)
35 / 36 Gwendal Simon Cloud Gaming
Main researchers in the areaKuan-Ta Chen (Sinica - Taiwan)
Shervin Shirmohammadi (Ottawa - Canada)
Mark Claypool (Worcester - UK)
Jiangchuan Liu (Simon Fraser - Canada)
Carsten Griwodz (Simula - Norway)
35 / 36 Gwendal Simon Cloud Gaming
Main researchers in the areaKuan-Ta Chen (Sinica - Taiwan)
Shervin Shirmohammadi (Ottawa - Canada)
Mark Claypool (Worcester - UK)
Jiangchuan Liu (Simon Fraser - Canada)
Carsten Griwodz (Simula - Norway)
35 / 36 Gwendal Simon Cloud Gaming
Main researchers in the areaKuan-Ta Chen (Sinica - Taiwan)
Shervin Shirmohammadi (Ottawa - Canada)
Mark Claypool (Worcester - UK)
Jiangchuan Liu (Simon Fraser - Canada)
Carsten Griwodz (Simula - Norway)
35 / 36 Gwendal Simon Cloud Gaming
Main researchers in the areaKuan-Ta Chen (Sinica - Taiwan)
Shervin Shirmohammadi (Ottawa - Canada)
Mark Claypool (Worcester - UK)
Jiangchuan Liu (Simon Fraser - Canada)
Carsten Griwodz (Simula - Norway)
35 / 36 Gwendal Simon Cloud Gaming
Takeaway
Business attractive, but techno challenging
Many research areas to explore :Delivery networkGame virtualizationPriority actions and adaptive encoding... and everything to chase the ms everywhere
And what about multi-player games ?
36 / 36 Gwendal Simon Cloud Gaming
Takeaway
Business attractive, but techno challenging
Many research areas to explore :Delivery network
Game virtualizationPriority actions and adaptive encoding... and everything to chase the ms everywhere
And what about multi-player games ?
36 / 36 Gwendal Simon Cloud Gaming
Takeaway
Business attractive, but techno challenging
Many research areas to explore :Delivery networkGame virtualization
Priority actions and adaptive encoding... and everything to chase the ms everywhere
And what about multi-player games ?
36 / 36 Gwendal Simon Cloud Gaming
Takeaway
Business attractive, but techno challenging
Many research areas to explore :Delivery networkGame virtualizationPriority actions and adaptive encoding
... and everything to chase the ms everywhere
And what about multi-player games ?
36 / 36 Gwendal Simon Cloud Gaming
Takeaway
Business attractive, but techno challenging
Many research areas to explore :Delivery networkGame virtualizationPriority actions and adaptive encoding... and everything to chase the ms everywhere
And what about multi-player games ?
36 / 36 Gwendal Simon Cloud Gaming