Software OpenFlow Solutions using Open vSwitch

Experimenter/Developer
Round Table:
Steroid OpenFlow Service
Ryan Izard
[email protected]
Steroid OpenFlow Service (SOS)
• Problem
– Large delay-bandwidth product networks
– TCP window size problem
– Cannot fill pipe
• SOS solution
–
–
–
–
Use multiple TCP connections
SDN-based
Transparent service improvement to user
Single and multipath support
SOS Design
• Agents
– Use multiple TCP connections
– Proxy for clients
• Clients and Servers
– Unmodified hosts; i.e. “end users”
– Unaware of SOS presence
• OpenFlow switches
– Perform packet redirection and rewrite
• Controller
– Detect, set up, and tear down SOS session
– Communicates with agents
General SOS Architecture
GENI Test Environment
Results in GENI
•
•
•
•
Without SOS: 15.4Mbps
With SOS:
85.6Mbps
Improvement: 5.6x
Notes and limitations
–
–
–
–
–
–
–
VMs
489MB RAM
Swap space used
32 parallel TCP connections
1410 byte MTU
GRE tunnel between sites
Disk-to-disk
AL2S Test Environment
AL2S Test Environment
•
•
•
•
10G end-to-end link b/t Clemson and Utah
Agent and client at each end
Floodlight controller at Clemson
Dell S4810 OpenFlow hardware switches at each site
Results in AL2S
•
•
•
•
Without SOS: 130Mbps
With SOS:
5.08Gbps (so far)
Improvement: ~40x
Notes and limitations
–
–
–
–
–
L3 and L4 packet rewrite in Open vSwitch at agents
L2 packet rewrite on Dell S4810s
7,000 parallel TCP connections
60,000 byte data chunk size
Clemson agent CPU usage near 100%  bottleneck
• Buy a better agent, or…
• Use multiple agents
Scalability in GENI
Scalability in GENI
Problems Encountered
GENI
• Stitching bandwidth
limitations
• Variable performance
over tunnel
(expected)
• Inconsistent VM
performance on
InstaGENI
Native AL2S
• VLAN translation
• OF instance config on
S4810s w/LLDP
• ARP on S4810s
• Inconsistent AL2S
performance
– Always better TX from
Clemson to Utah
– Also true for CloudLab
Possible Solutions
GENI
• Stitching bandwidth
limitations  ?
• Variable performance
over tunnel
(expected)
• Inconsistent VM
performance on
InstaGENI  more
fine grained
monitoring
Native AL2S
• VLAN translation
• OF instance config on
S4810s w/LLDP
• ARP on S4810s
• Inconsistent AL2S
performance
– Always better TX from
Clemson to Utah
– Also true for CloudLab
–  monitoring of AL2S