1. NetCloud Hong Kong
2017/12/11
NetCloud Hong Kong
2017/12/11
PA-Flow:
Gradual Packet Aggregation at Virtual Network I/O for Efficient Service Chaining
Yuki Taguchi†, Ryota Kawashima†, Hiroki Nakayama††,
Tsunemasa Hayashi††, and Hiroshi Matsuo†
† Nagoya Institute of Technology, Japan
†† BOSCO Technologies Inc.

2. Boosting per-flow performance of service chains
The Goal
Boosting per-flow performance of service chains
VNF
NFV-node
VM or Container
Forwarding performance of NFV-nodes is poor.
Virtual Network Function
VNF1
VNF2
VNF4
VNF3
VNF1
PA-Flow
VNF3
PA-Flow
VNF3
VNF4
PA-Flow
VNF2
PA-Flow
overloaded

3. Contents Background Related Work Proposal Evaluation Conclusion1
Background
Problems of Service Chaining
Related Work
Scaling-out
Scaling-up
Proposal
Evaluation
Conclusion 2 3 4 5

4. Current Status of NFV Softwarized devices degrade the performance
NFV services have been launched
Service chaining is a key concept of NFV
Toy-blocking style service chain
vFirewall
vDPI
vGateway
Softwarization of devices
Softwarized devices degrade the performance
Network-Level problems
Node-Level problems

5. Network-Level Problems
Server
1st VNF
bottleneck
2nd VNF
Server
3rd VNF
Server
Server
Server
Another network
Server
Server
Server
upper-stream
The performance of the upper VNFs must be ensured

6. Node-Level Problems Performance degradation factorsVM
VNF
Bottleneck!
vNetwork I/O
Host
Bottleneck!
Dropped!
vSwitch
too many packets
NFV-node
Performance degradation factors
Flow matching by virtual switches
Packet copy and queueing at vNetwork I/O

7. Performance of NFV-nodes
Huge gap !
Small gap
“Enlarging packet size” is the key!
† R. Kawashima et al., “Evaluation of Forwarding Efficiency in NFV-nodes toward Predictable Service
Chain Performance,” IEEE Trans. on Network and Service Management, 2017.

8. Contents Background Related Work Proposal Evaluation Conclusion1
Background
Problems of Service Chaining
Related Work
Scaling-out
Scaling-up
Proposal
Evaluation
Conclusion 2 3 4 5

9. Related Work Scaling-Out Scaling-Up 1. Auto Scaling 2. NIC Offloading
vSwitch
SmartNIC
(FPGA inside)
On-demand instantiation
VNF
VNF
VNF
3. Packet Aggregation
Small packets
OpenFlow SW
A single large packet

10. 1. Auto-Scaling Physical resources are limited
new instance
VNF
VNF
overloaded!
VNF
healthy
healthy
healthy
instantiation
SDN Controller
flow rule
Physical resources are limited
There is no de-facto framework for monitoring
† Shoumik Palkar et al., "E2: A Framework for NFV Applications"
ACM Symposium on Operating Systems Principles 2015 (SOSP' 15)

11. Low Latency & High Throughput
2. NIC Offloading
Smart NIC†
Host
Offloading Flow Table
Low Latency & High Throughput
Smart NIC with FPGA
It only works with a specific vSwitch (e.g., Open vSwitch)
Performance gain is limited
† Agilio OVS Software, combind with Agilio SmartNICs

12. 3. Aggregation Approach Aggregation between VMs and the Host† VM Host
VNF
Extra queueing
latency is added VM
Aggregation
Packet Batch
Packet Queue
No. outgoing packets
is not reduced
Host
Disaggregation
DPDK is not supported
† M. Bourguiba et al., “Improving Network I/O Virtualization for Cloud Computing“
IEEE Trans. on Parallel and Distributed Systems, 2014

13. Contents Background Related Work Proposal Evaluation Conclusion1
Background
Problems of Service Chaining
Related Work
Scaling-out
Scaling-up
Proposal
Evaluation
Conclusion 2 3 4 5

14. (Packet Aggregation Flow)
Proposal
PA-Flow
(Packet Aggregation Flow)
Key points
1. Fast and Network-aware Aggregation
2. Gradual Packet Aggregation
3. Next-hop-aware Aggregation

15. 1. Fast and Network-aware Aggregation
VNF
No extra queueing latency! VM
PA-Flow
 explained later
PA-Flow
header
Outgoing packets are reduced
Host
PA-Flow
header
VXLAN
DPDK is supported
All the existing problems are solved!

16. 2. Gradual Packet Aggregation
Gradually aggregated
Free space
PA-Flow
enabled VM
Host
PA-Flow
Full MTU size
PA-Flow
enabled
The average packet size is enlarged!

17. 3. Next-hop Aware Aggregation
SDN Controller
Flow Table
Identifier
Next-Hop 3
100 5
200
Next-Hop
(ID: 100) VM
Host
PA-Flow
branch 3
Next-Hop
(ID: 200)
Add/Delete Entries 5
PA-Flow supports branches of service chain

18. Overview Architecture
: disaggregation
: aggregation
2. Gradual aggregation
PA-Flow
Firewall
DPI
Policer
Firewall
3. Next-hop-aware
Gateway
DPI
Firewall
1. Fast and Network-aware aggregation

19. PA-Flow enabled NFV-nodeVM
VNF
Packets sent by VNF
packet queue
Host
PA-Flow
header
Flow Table
Same Next-hop
Disaggregation
Header format
version
(8 bit)
Packet
length 1
(16 bit)
type
packet num
packet
length 2 …
Aggregation
PA-Flow
VXLAN
header
MAC IP
UDP
PA-Flow
header
vSwitch
vSwitch processing (VXLAN) can be used

20. There is no extra waiting time for aggregation
Implementation
(a) Default vhost-user
(b) PA-Flow
VNF
VNF
Enqueue at once
packet queue
packet queue
polling … …
header
Copy at once
packet structure
packet structure
There is no extra waiting time for aggregation

21. Contents Background Related Work Proposal Evaluation Conclusion1
Background
Problems of Service Chaining
Related Work
Scaling-out
Scaling-up
Proposal
Evaluation
Conclusion 2 3 4 5

22. Performance Evaluation
Exp. 1: Baseline
CPU
i7-6900K
(8 cores, 3.2 GHz)　
NIC
10 GbE (Intel X540) OS
CentOS 7.3
Tester
NFV-node
DuT
Exp. 2: Service Chain
VXLAN used
Tester …
NFV-node
Tester
NFV-node
More realistic
NFV-system
Tester
2-VNF chain

23. Exp.1: Baseline (a) Default (b) PA-Flow (c) PA-Flow with VNF support
VM A
VM B
Traffic Generator
(MoonGen)
VNF
(OVS-DPDK)
vport 3
vport 4
vport 3
vport 4
Host A
Host B
OVS-DPDK
OVS-DPDK
port 1
port 2
port 1
port 2
DuT
10Gb Ethernet

24. Exp.1: Baseline (a) Default (b) PA-Flow (c) PA-Flow with VNF support
VM A
VM B
Traffic Generator
(MoonGen)
VNF
(OVS-DPDK)
PA-Flow
vport 3
vport 4
vport 3
vport 4
Host A
Host B
Disaggr.
Aggr.
Disaggr.
Aggr.
OVS-DPDK
OVS-DPDK
port 1
port 2
port 1
port 2
DuT
10Gb Ethernet

25. Exp.1: Baseline (a) Default (b) PA-Flow (c) PA-Flow with VNF support
VM A
VM B
Traffic Generator
(MoonGen)
VNF
(OVS-DPDK)
PA-Flow
vport 3
vport 4
vport 3
vport 4
This VNF handles
PA-Flow packet
PA-Flow
header
Host A
Host B
Disaggr.
Aggr.
OVS-DPDK
OVS-DPDK
port 1
port 2
port 1
port 2
DuT
10Gb Ethernet

26. The performance is improved to nearly 10 Gbps
Exp.1: Throughput
175%
75%
The performance is improved to nearly 10 Gbps

27. Exp.1: Latency / Jitter 200 ns increased 2 µs reduced
Tx rate: 1 Mpps
(1% aggregated)
Tx rate: 3 Mpps
(25% aggregated)
200 ns increased
2 µs reduced
The pure overhead is negligible
The latency reduced under high-rate communications

28. Exp.2: Service Chain 2-VNF chain Tester1 Tester3 VNF1 VNF2 Tester2
Double Tx rate
Tester1
host1
Traffic Generator
( Tx only )
Port
Tester3
host3
Port
Traffic Generator
( Tx only )
VXLAN used
host4
VNF1
Port
OVS-DPDK
VNF2
host5
Port
OVS-DPDK
Tester2
host2
Traffic Generator
( Tx and Rx)
Port
2-VNF chain

29. The performance of service chain is drastically boosted
Exp.2: Throughput
370%
Overloaded at 3.2 Mpps
170%
Ideal throughput!
The performance of service chain is drastically boosted

30. Contents Background Related Work Proposal Evaluation Conclusion1
Background
Problems of Service Chaining
Related Work
Scaling-out
Scaling-up
Proposal
Evaluation
Conclusion 2 3 4 5

31. Conclusion Performance limitation of Service Chain
Software-based scale-up approach is needed.
Proposal: PA-Flow (Packet Aggregation Flow)
Gradual aggregation reduces forwarding cost of upper VNFs.
Flexible operation is realized with next-hop aware aggregation.
Performance improved by 170% on a 2-VNF service chain.
Future work
Further performance improvement
Porting PA-Flow to a VM-side component
Evaluation of container