The session provides an update to on the DPACC project within the OPNFV with a brief discussion on APIs and implementation progress. This session will review the API definition progress and follow up with a demo highlighting a common application as the vNF running on top of the DPACC defined layers. The demo will highlight the use of both hardware and software acceleration utilizing the DPACC defined acceleration layers. The demonstrationIt will highlight the progress in optimizing performance and latency characteristics of a platform to realize the vision of NFV while meeting stringent requirements, particularly for certain workloads, required by carriers.

1. DPACC Acceleration
Progress and Demonstration
DPDK on Axxia ARM hardware for DPACC
Keith Wiles, Intel Corp
Magnus Karlsson, Intel Corp

2. Demo of DPDK on ARM
Using Axxia ARM hardware with
DPDK to get the best performance

3. What is the demonstration?
• The goal of the demo is to show DPDK running on ARM
hardware with good performance
• We have ported DPDK to the Axxia ARMv7 hardware
and integrated the Axxia SoC hardware
• The demo shows that DPDK is similar in performance to
the native SoC environment
10/30/15 DPACC Acceleration demonstrationwith DPDK on ARM 3

4. DPDK ported to ARM, why?
• DPDK was ported to ARM hardware to demonstrate how
portable DPDK is to any platform
• The port only took a few days to get running on the Axxia
platform (not optimized yet)
• Next few days we integrated the Axxia SoC drivers
• Demonstrating DPDK was easy and quick to run on any
platform
• Plus DPDK on ARM hardware shows performance similar to
the native SoC application
10/30/15 DPACC Acceleration demonstrationwith DPDK on ARM 4

5. How the port of DPDK was done
• DPDK contains infrastructure to add different platforms in the EAL layer
– A new platform was created for ARMv7 that used generic C built-ins or ARM assembly code
to support platform specific functions
• DPDK also supports virtual PCAP based PMDs and test applications
– Using them, the port could be validated easily and quickly
• Addition of a virtual PMD is straight forward for DPDK
– A new virtual PMD was added to create Ethernet devices abstracting Axxia ports
– DPDK’s APIs could be supported by using Axxia specific SDK calls underneath
• No change was done in any sample/test application of DPDK to run them using the new Axxia
PMD
10/30/15 DPACC Acceleration demonstrationwith DPDK on ARM 5

6. 6
DPDK on Axxia Environment
Ethernet Switch
Axxia_Eth0
00:00:55:16:00:00
Axxia_Eth1
00:00:55:16:00:01
DPDK
l2fwd
Linux Machine Axxia Hardware
Data traffic
Axxia
l2fwdDPDK packet generator
application to send and
receive packets of two
flows on one port to
measure latencies
Eth1
02:00:00:00:00:01
Flow 0 – Traverses via DPDK
SMAC – 02:00:00:00:00:01
DMAC – 00:00:55:16:00:01
SMAC – 00:00:55:16:00:00
DMAC – 02:00:00:00:00:01
SMAC – 00:00:55:16:00:01
DMAC – 02:00:00:00:00:01
SMAC – 02:00:00:00:00:01
DMAC – 00:00:55:16:00:00
Flow 1 – Traverses via Axxia SDK
P0 P1 P2
10/30/15 DPACC Acceleration demonstrationwith DPDK on ARM 6
DPDK based packet generator and receiver
applications, which also computes latency

7. Graph of DPDK compared to Axxia Native SoC software
10/30/15 DPACC Acceleration demonstrationwith DPDK on ARM 7

8. Summary
• The goal was to show DPDK does not add any overhead in
performance compared to a native acceleration SDK
• We created a ARMv7 port, which we have pushed up stream to
DPDK.org
– The port needs to be optimized,but the port shows we still
obtain good performance compared to a native design
• The effort for the ARMv7 port and the comparison of two
systems was a reasonable amount of time (couple of months)
• Integrating to ARM and a vendor SDK for DPDK provides good
performance and minimum overhead
10/30/15 DPACC Acceleration demonstrationwith DPDK on ARM 8

9. DPDK on ARM for DAPCC
Questions?
Answers: 42 or 47 take your pick J

10. Thank you for Attending!