Hardware Acceleration of a Software-based VPN Furkan Turan Ruan de - - PowerPoint PPT Presentation

Hardware Acceleration

• f a Software-based VPN

Furkan Turan Ruan de Clercq, Pieter Maene, Oscar Reparaz Ingrid Verbauwhede KU Leuven - COSIC

VPN Introduction

2

VPN (Virtual Private Network) encrypts the communication between two parties.

VPN Device Introduction

Goal: Start with a VPN application, Convert it into a 2 port VPN device, Accelerate it with a cryptographic coprocessor.

VPN Device

VPN Device

3

Software-based VPN

How a software-based VPN application works: SigmaVPN: Light-weight, secure and modular software-based VPN

Application Application Virtual Network Interface VPN Application Physical Network Interface

4

2 Port VPN Device with Hardware Accelerator

Private Comm Public Comm Sigma VPN Coprocessor Linux

The new Private Comm. module uses a Physical Network Interface. It is capable of even capturing broadcast messages.

5

NaCl’s CryptoBox

KS ← ECDH(KSEC,A, KPUB,B) KS ← ECDH(KSEC,B, KPUB,A) KD ← HSalsa20(KS, N1) S ← Salsa20(KD, N2 || CTR) CT ← S ⊕ MSG MACA ← Poly1305(CT, S) KD ← HSalsa20(KS, N1) S ← Salsa20(KD, N2 || CTR) MACB ← Poly1305(CT, S) Compare(MACA, MACB) MSG ← S ⊕ CT Alice Bob CT, MACA, N1,2

7

One-time Authenticator: Poly1305

Msg[0:127] Acc + x

mod 2130-5

R Acc Msg[128:256] + x

mod 2130-5

Msg[x:x+128] ... Acc + x

mod 2130-5

Acc Acc S +

mod 2128

MAC

128 130 128 131 128

…

An update operation for each 128-bit blocks of the message The operation implements a modular multiplication in radix (2130-5)

9

Poly1305’s Implementation

Implemented using a school-book multiplication:

• Big multiplication is divided into smaller blocks
• Followed by propagation of the results

Each small block multiplication is handled in single-cycle multipliers of Zynq’s DSP48 Slices To boost the performance:

• Parallel execution of smaller-block multiplications
• Parallel propagating the results

x5 x5

10

Poly1305’s Implementation

11

A datapath for each column to handle smaller block multiplications. Result of a column is propagated to the next. The multipliers set the critical path.

Block of Operand 1 Block Operand 2 X x5 +

Hardware Implementation

• Processing System runs Linux - SigmaVPN.
• DMA transfers data between co-processor and RAM.

ZYNQ Processing System (PS) ARM Cores Programmable Logic (PL) DMA

CTRL Dest Add. Length

Cryptographic Coprocessor

AXI4 Lite AXI4 Full AXI4 Stream

12

Coprocessor's Datapath

512 512 512 512 128

13

Scheduling

• Operation is divided into time slots
• A time slot is the time to process a 512-bit message block
• Each hardware module is active in each time slot

14

Hardware Utilization

Single Instance of Processing Blocks

• Resource Utilization: 53.67%
• Max Clock Freq: 92.85 MHz
• Process 512-bit block in a time slot

Duplicated Processing Blocks

• Resource Utilization: 97.25%
• Max Clock Freq: 81.25 MHz
• Process 1024-bit block in a time slot

15

ZYBO Board comes with Zynq Z-7010 SoC;

• The smallest Zynq device
• Has limited resources

Communication btw. HW & SW

Configuring DMA for transferring buffers requires:

• Accessing physical addresses
• Coherent memory accesses

Created a Linux kernel space module (Device File) Problem: Overhead of making context switches

• Going do kernel space costs ~800 cycles.
• Transferring the frame btw. User and Kernel space costs ~740 cycles.

16

Improvements to Cryptographic Operations

• Encrypted and decrypted many test vectors with

both SW-only and SW+HW implementations.

• Compared results for accuracy and execution times.

2 4 6 8 10 12 14 16 18 16 32 64 128 256 512 1024

Improvement (Factor) Message Length (Bytes)

Encryption Decryption

Improvement in Encryption Min 4.9, Max 15.1 Improvement in Decryption Min 9.1, Max 16.2

17

Improvements to VPN Bandwidth

Test Network Structure: Bandwidth tests using Iperf Network Bandwidth Measurement Tool

VPN Device VPN Device

18

Improvements to VPN Bandwidth

TCP bandwidth increase

• 2.9 times for 128-byte frames,
• 4.36 times for 1024-byte frames.

UDP bandwidth increase

• 2 times for 128-byte frames,
• 5.36 times for 1024-byte frames.

10 20 30 40 50 60 70 80 90 100 TCP UDP

Bandwidth (Mbps) for Comm. with 1024-byte ETH Frames

No VPN VPN without Crypto VPN with SW Crypto VPN with HW+SW Crypto

19

Functionality Test

• The designed VPN device is still capable of

establishing a secure communication with

• riginal SigmaVPN application.
• A VPN device on a low-cost dev-board,

providing confidential communication between a whole home/business network and a remote server.

20

Conclusion

• A cryptographic hardware accelerator is offered for

NaCl's CryptoBox specifically for SigmaVPN.

• Encrypting a 1024-byte message in 94% less time

compared to SW-only implementation.

• Integrating our HW-SW codesign into SigmaVPN
• ffers up to 6 times more communication bandwidth.
• Xilinx Open HW Design Contest Finalist:

http://www.openhw.eu/2016-finalists.html

• It’s available open source:

https://github.com/furkanturan/Hardware-Accelerated-SigmaVPN

21