Timing Analysis of Linux-Based CAN-to-CAN Gateway Michal Sojka - PowerPoint PPT Presentation

Timing Analysis of Linux-Based CAN-to-CAN Gateway Michal Sojka Czech Technical University in Prague Faculty of Electrical Engineering sojkam1@fel.cvut.cz February 3, 2011 Center for Applied Cybernetics Embedded Systems Colloquium Michal Sojka Timing Analysis of Linux-Based CAN-to-CAN Gateway February 3, 2011 1 / 21

Outline Introduction Testbed Description Results Conclusion Michal Sojka Timing Analysis of Linux-Based CAN-to-CAN Gateway February 3, 2011 2 / 21

Introduction SocketCAN ◮ Open-source project, official CAN-bus support for Linux kernel ◮ Initiated by Volkswagen research – first open source project where Volkswagen participates ◮ Core developers also from a few other companies ◮ Contributors from many other people ◮ In mainline Linux since 2.6.25 ◮ Based on standard Linux networking infrastructure ◮ http://socketcan.berlios.de Michal Sojka Timing Analysis of Linux-Based CAN-to-CAN Gateway February 3, 2011 3 / 21

Introduction SocketCAN-based Gateway ◮ Routes CAN frames between multiple CAN buses. ◮ Frames can be modified (ID, length, data) ◮ Modifications: SET, AND, OR, XOR, CRC ◮ Can filter frames based on ID/mask pairs. ◮ Implemented as Linux kernel module ◮ Easy to use Michal Sojka Timing Analysis of Linux-Based CAN-to-CAN Gateway February 3, 2011 4 / 21

Introduction The Goal of our Work ◮ Timing analysis of the gateway ◮ How are the routed frames delayed by the gateway? ◮ What influences the delays (latencies)? ◮ Traffic patterns, ◮ additional load, ◮ gateway configuration, ◮ Linux kernel version. ◮ Volkswagen uses the gateway for fast prototyping. It works well for them and they wanted to know where are the limits. Michal Sojka Timing Analysis of Linux-Based CAN-to-CAN Gateway February 3, 2011 5 / 21

Testbed Description Testbed Setup ◮ Gateway – PowerPC MPC5200 ◮ PC – Pentium 4, Kvaser PCI quad-CAN SJA1000-based Michal Sojka Timing Analysis of Linux-Based CAN-to-CAN Gateway February 3, 2011 6 / 21

Testbed Description Measurement Methodology ◮ Bus bit-rate 1 Mbps ◮ Frames are sent from can0 ◮ Received on can1 and can2 ◮ Receive timestamps are taken by the driver at interrupt time – SO_TIMESTAMPNS socket option ◮ Timestamps with nanosecond precision (TSC) ◮ Same clock for both timestamps – no problems with time synchronization between computers msg 1 CAN bus 0 CAN bus 1 msg 1' time message latency Michal Sojka Timing Analysis of Linux-Based CAN-to-CAN Gateway February 3, 2011 7 / 21

Testbed Description Measurement Measurement II. ◮ Traffic patterns: ◮ one message at a time, ◮ 50% bus load, ◮ flood (almost 100% bus load) ◮ Additional loads: ◮ CPU load – hackbench ◮ Ethernet load – ping -f -s 6000 ◮ Versions: ◮ Latest SocketCAN from SVN (rev 1199 + fixes) ◮ Kernel: 2.6.33, 2.6.33-rt, 2.6.36 Michal Sojka Timing Analysis of Linux-Based CAN-to-CAN Gateway February 3, 2011 8 / 21

Results Precision and Presentation of Results No GW, two interfaces in PC (GW kernel 2.6.33.7-00007-g9c0ff90, Traffic 50, Load none) 10000 2 byte messages 4 byte messages Latency profile [messages] 6 byte messages 8 byte messages 1000 100 10 1 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 Time [ms] ◮ Two points ( t 1 , m 1 ) and ( t 2 , m 2 ) : t 1 < t 2 ⇒ m 1 − m 2 frames had the latency in the range [ t 1 , t 2 ) . ◮ The right most point ( t w , m w ) : m w frames with the latency of t w . Michal Sojka Timing Analysis of Linux-Based CAN-to-CAN Gateway February 3, 2011 9 / 21

Results Simple Gateway Single GW job for all messages, no modifications Single GW job for all messages, no modifications (GW kernel 2.6.33.7-00007-g9c0ff90, Traffic flood, Load none) (GW kernel 2.6.33.7-00007-g9c0ff90, Traffic flood, Load none) 10000 0.3 2 byte messages 2 byte messages 4 byte messages 4 byte messages Latency profile [messages] 6 byte messages 6 byte messages 0.25 8 byte messages 8 byte messages 1000 Latency [ms] 0.2 100 0.15 10 0.1 1 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0 1 2 3 4 5 6 7 Time [ms] Time [s] Michal Sojka Timing Analysis of Linux-Based CAN-to-CAN Gateway February 3, 2011 10 / 21

Results Frame Filtering 2048 GW jobs (one per id, mask C00007FF), no modifications Two types of filters (GW kernel 2.6.33.7-00007-g9c0ff90, Traffic oneatatime, Load none) 10000 Message id 0 ◮ SFF frames (11 bit ID), Message id 255 Latency profile [messages] Message id 511 Message id 767 1000 single ID Message id 1023 Message id 1279 Message id 1535 Message id 1791 ◮ Implemented as lookup 100 Message id 2047 table (by ID) 10 ◮ Fast 1 0.05 0.1 0.15 0.2 0.25 0.3 ◮ Top figure Time [ms] ◮ EFF frames (29 bit ID) or 2048 GW jobs (one per id, mask 0x7FF), no modifications (GW kernel 2.6.33.7-00007-g9c0ff90, Traffic oneatatime, Load none) 10000 filters matching multiple Message id 0 Message id 255 Latency profile [messages] Message id 511 frames (mask) Message id 767 1000 Message id 1023 Message id 1279 ◮ Implemented as linear Message id 1535 Message id 1791 100 Message id 2047 list 10 ◮ Latency depends on ID ◮ Bottom figure 1 0.4 0.45 0.5 0.55 0.6 0.65 0.7 0.75 0.8 Time [ms] Michal Sojka Timing Analysis of Linux-Based CAN-to-CAN Gateway February 3, 2011 11 / 21

Results Frame Filtering under Heavy Traffic 2048 GW jobs (one per id, mask 0x7FF), no modifications (GW kernel 2.6.33.7-00007-g9c0ff90, Traffic flood, Load none) 10000 Message id 0 Message id 255 Latency profile [messages] Message id 511 1000 Message id 767 Message id 1023 Message id 1279 Message id 1535 100 Message id 1791 ◮ Packet processing takes Message id 2047 10 longer time ⇒ packet losses 1 0.1 1 10 100 Time [ms] ◮ The effect of hardware 2048 GW jobs (one per id, mask 0x7FF), no modifications (GW kernel 2.6.33.7-00007-g9c0ff90, Traffic flood, Load none) buffers (bottom figure) 100 Message id 0 Message id 255 Message id 511 ◮ MPC5200 has 4 hardware Message id 767 10 Message id 1023 Latency [ms] Message id 1279 Message id 1535 buffers Message id 1791 Message id 2047 1 0.1 0 2 4 6 8 10 12 14 16 18 20 Time [s] Michal Sojka Timing Analysis of Linux-Based CAN-to-CAN Gateway February 3, 2011 12 / 21

Results Frame Modification Single GW job for all messages with modifications, 8 byte messages (GW kernel 2.6.33.7-00007-g9c0ff90, Traffic oneatatime, Load none) 10000 Latency profile [messages] 1000 100 No modifications 10 T wo modifications Four modifications Four modifications and XOR checksum Four modifications and CRC8 checksum 1 140 150 160 170 180 190 200 Time [µs] ◮ Perpetum mobile – the more work to do, the less time it takes. ◮ Wrong branch prediction? No. Difference under 1 µ s . Michal Sojka Timing Analysis of Linux-Based CAN-to-CAN Gateway February 3, 2011 13 / 21

Results Multi-Hop Routing with Virtual Interfaces MPC5200 GW0 GW(n+1) id id id+n id+n can0 vcan0 can1 nop nop id id+n id+1 id+2 id+1 i d + n - 1 GW1 GW2 GWn id ⇒ id+1 id ⇒ id+1 id ⇒ id+1 Figure: Multi-Hop gateway with a single virtual CAN interface (vcan0). MPC5200 GW0 GW1 GW2 id+1 id+2 id id+1 id+2 id+3 can0 vcan0 vcan1 can1 id ⇒ id+1 id ⇒ id+1 id ⇒ id+1 Figure: Multi-Hop gateway with several virtual CAN interfaces. Michal Sojka Timing Analysis of Linux-Based CAN-to-CAN Gateway February 3, 2011 14 / 21

Results Multi-Hop Routing with Virtual Interfaces Chained GW jobs on a single VCAN interface Chained GW jobs on multiple VCAN interfaces (GW kernel 2.6.33.7-00007-g9c0ff90, Traffic oneatatime, Load none) (GW kernel 2.6.33.7-00007-g9c0ff90, Traffic oneatatime, Load none) 10000 10000 2 GW jobs, 1 vcan 1 GW jobs, 0 vcans 4 GW jobs, 1 vcan 2 GW jobs, 1 vcans Latency profile [messages] Latency profile [messages] 6 GW jobs, 1 vcan 4 GW jobs, 3 vcans 1000 8 GW jobs, 1 vcan 1000 6 GW jobs, 5 vcans 10 GW jobs, 1 vcan 8 GW jobs, 7 vcans 12 GW jobs, 1 vcan 10 GW jobs, 9 vcans 100 100 10 10 1 1 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.1 0.15 0.2 0.25 0.3 0.35 0.4 Time [ms] Time [ms] ◮ Goal: evaluate the internals of the gateway and SocketCAN infrastructure. ◮ Result: nothing unexpected – the more hops, the more time. ◮ Every additional hop adds approximately 11 µ s . Michal Sojka Timing Analysis of Linux-Based CAN-to-CAN Gateway February 3, 2011 15 / 21

Results Kernel-Space vs. User-Space routing – Non-rt Kernel Kernel vs. userspace GW, no modifications Kernel vs. userspace GW, no modifications (GW kernel 2.6.33.7-00007-g9c0ff90, Traffic oneatatime, Load none) (GW kernel 2.6.33.7-00007-g9c0ff90, Traffic flood, Load none) 10000 10000 Userspace GW, 2 byte messages Userspace GW, 2 byte messages Userspace GW, 8 byte messages Userspace GW, 8 byte messages Latency profile [messages] Latency profile [messages] Kernel GW, 2 byte messages Kernel GW, 2 byte messages Kernel GW, 8 byte messages Kernel GW, 8 byte messages 1000 1000 100 100 10 10 1 1 0.1 1 10 100 1000 0.1 1 10 100 1000 Time [ms] Time [ms] ◮ Kernel-space has much lower overhead. ◮ Kernel-space handles even flood traffic without problems. Michal Sojka Timing Analysis of Linux-Based CAN-to-CAN Gateway February 3, 2011 16 / 21