1. QoE-Aware Traffic Steering using
OpenFlow
Prasad Calyam, Ph.D.
US Ignite and ONF Workshop,
October 8th 2013
Research Sponsors: NSF (CNS-1050225, CNS-1205658), VMware, Dell, IBM
http://vmlab.oar.net
2. Discussion Topics
• User QoE Problem Context
• Solution Approach and Results
• “One more thing…”
2
3. Discussion Topics
• User QoE Problem Context
• Solution Approach and Results
• “One more thing…”
3
4. Virtual Desktop Clouds (DaaS)
“Brain of the Cloud”
P. Calyam, R. Patali, A. Berryman, A. Lai, R. Ramnath, “Utility-directed Resource Allocation in Virtual Desktop Clouds”, Elsevier
Computer Networks Journal (COMNET), 2011. 4
5. Example DaaS Use Cases
(a) Virtual classroom lab involving faculty and students
(b) Computationally intensive interactive applications for biomedical community
(e.g., remote volume visualization)
(c) Simulation-as-a-Service requiring HPC resources for advanced manufacturing
(d) Virtual desktops for underserved communities
5
P. Calyam, A. Berryman, A Lai, M. Honigford, “VMLab: Infrastructure to Support Desktop Virtualization Experiments for Research and
Education”, VMware Technical Journal (Invited Paper), 2012.
6. Research Scientist
Home User
Mobile User
Fixed Resource
Allocation Model
• High consistent CPU
• High consistent memory
• High bandwidth connectivity
• Low bursty CPU
• Low bursty memory
• Medium bandwidth connectivity
• Low bursty CPU
• Low bursty memory
• Low bandwidth connectivity
CPU
Memory
Bandwidth
VDCs Today – Overprovisioning and Guesswork…
Available Resources
Number of Users
VDC Service Provider
Unified
Resource Broker
=
6
7. Overprovisioning and Guesswork Fails!
Home User
Mobile User
VDC Service Provider
• Inadequate CPU, memory and bandwidth
(Impact e.g., Slow interaction response times)
• Calls from unhappy customers
• High operation $$
Problem: Resource allocation without
awareness of system, network and
user experience characteristics
• Inadequate CPU, memory and bandwidth
(Impact e.g., IPTV with impairments and slow playback)
• Excess CPU, memory and bandwidth
(Impact e.g., Good interaction response times and
smooth IPTV playback)
Research Scientist
7
8. VDCs in the Future – Smart thin-clients at user sites
Smart
Thin-Client
Smart
Thin-Client
VDC Service Provider
• Happy customers
• Low operation $$
Research Scientist
Home User
Mobile User
CPU
Memory
Bandwidth
• Utility-directed CPU, memory and bandwidth
(Impact e.g., Good interaction response times and
smooth IPTV playback)
Unified
Resource Broker
Utility-directed
Dynamic Resource
Allocation Model
(U-RAM)
=
8
NOTE: Application behavior profiles collected from smart thin-client feedback also help
in QoE degradation troubleshooting!**
** Y. Xu, P. Calyam, D. Welling, S. Mohan, A. Berryman, R. Ramnath, “Human-centric
Composite Quality Modeling and Assessment for Virtual Desktop Clouds”, ZTE
Communications Journal (Invited Paper), 2013.
9. VD Placement after U-RAM Provisioning
• URB Placement decisions involving data centers are influenced by:
– Session latency, Load balancing, Operation cost
• Placement decisions need to be changed over time -
– Proactive Defragmentation for improved performance and scalability
• Opportunistic placement reduces user wait time for access initially, but over
time causes resource fragmentation due to changing application workloads
– Resource fragmentation decreases scalability (VDs/core) and
performance (user QoE), hence the VDC Net-Utility
» Net-Utility is a overall user QoE measurement across the VDC
– Reactive Migrations for increased resilience and sustained availability
• Cyber-attacks or planned maintenance necessitate VD migrations without
drastically affecting VDC Net-Utility
• We have developed proactive and reactive placement schemes
9
M. Sridharan, P. Calyam, A. Venkataraman, A. Berryman, “Defragmentation of Resources in Virtual Desktop Clouds for Cost-Aware
Utility-Optimal Allocation”, IEEE Conf. on Utility and Cloud Computing (UCC), 2011.
10. Problem Context Summary
• To use OpenFlow for dynamic resource placement of VD
applications via an URB and accomplish:
– Provisioning of non-IP VD application traffic flows between thin-
client sites and data centers based on utility functions
– Path selection and load-balancing of VD flows to ensure
satisfactory user QoE of interactive applications (e.g., video playback)
– Leveraging in-band instrumentation and measurement to gather
performance intelligence on cross traffic impact affecting VD
– Automated management and centralized network as well as
measurement control
10
11. Discussion Topics
• User QoE Problem Context
• Solution Approach and Results
• “One more thing…”
11
12. VIMAN Lab’s “VDC-Analyst”
VD Provisioning and Placement
GENI Slice Testbed for VDC Hosting
• VDC-Analyst → GENI
• Design & Development →
Validation and design tuning
• Large-scale simulations →
Cloud deployment experiments
12
13. VDC Architecture
Data Center OpenFlow Switches Thin-clients
Unified Resource Broker
Connection
Broker
Marker Packet
Handler
Packet
Capture
OpenFlow Switch
Flow tables Group Tables
Data Plane
Packet/Flow
Inspector
Routing Engine
Thin-client
Virtual Desktop
Secure
Channel
User Applications
Hypervisor
Security Token
RDP/PCoIP Server
Active
Directory
RDP/PCoIP Client
Load
Balancing
Control
Plane
Service Engine
Measurement
Plane
System
Provisioning
File System
Resource
Optimization
Secure
Channel
Control
Plane
OpenFlow
Controller
Measurement Engine
Active
Measurement
Congestion
Detection
Fault
Detection
13
P. Calyam, S. Rajagopalan, A. Selvadhurai, S. Mohan, A. Venkataraman, A. Berryman, R. Ramnath, “Leveraging OpenFlow for Resource
Placement of Virtual Desktop Cloud Applications”, IFIP/IEEE International Symposium on Integrated Network Management (IM), 2013 .
17. OpenFlow
Switch
Client In
Port
Out
Port
SUNNW PG48 50 51
SUNNW PG49 50 51
ATLANTA PG46 52 52
ATLANTA PG47 52 52
ATLANTA PG46 20 52
ATLANTA PG47 20 52
VDC-Analyst OpenFlow Demonstration
Route setupStep-1
Cross-traffic
Impact
Step-2
Load-balancing
ImprovementStep-3
OpenFlow
Switch
Client In
Port
Out
Port
ATLA PG46 20 52
ATLA PG47 20 52
OpenFlow
Switch
Client In
Port
Out
Port
ATLANTA PG46 20 52
ATLANTA PG47 20 52
SUNNW PG48 50 52
SUNNW PG49 50 52
Video runs smooth, GUI
applications are responsive
Video freezes, disconnects, GUI
applications are not responsive
Video runs smooth, GUI
applications are responsive
17
18. 0.21
15.36
0
5
10
15
20
Application Cross-Traffic
VDC-Analyst OpenFlow Demonstration
Route setupStep-1
Cross-traffic
Impact
Step-2
Load-balancing
ImprovementStep-3
Video runs smooth, GUI
applications are responsive
Video freezes, disconnects, GUI
applications are not responsive
Video runs smooth, GUI
applications are responsive
Bandwidth Consumed (Mbytes/s)
4.45
14.8
0
5
10
15
20
Application Cross-Traffic
4.6
0
0
5
10
15
20
Application Cross-Traffic
18
19. Discussion Topics
• User QoE Problem Context
• Solution Approach and Results
• “One more thing…”
19
20. User QoE Degradation Troubleshooting
• End-to-end user QoE degradation troubleshooting with OpenFlow
over multi-domain Layer 2 networks
20
Slow-motion benchmarking
of thin-client performance –
VDBench Tool
Real-time Capture and Analysis of
Packet Traces of User Tasks
(without using spanning ports)