Insights into the performance and configuration of TCP in Automotive Ethernet Networks

(1)

Insights into the performance and

configuration of

TCP in Automotive Ethernet Networks

Jörn MIGGE, RealTime-at-Work (RTaW)

Nicolas NAVET, University of Luxembourg

2018 IEEE Standards Association (IEEE-SA) Ethernet & IP @ Automotive Technology Day

9-10 October 2018 | London, England

(2)

Use-cases for TCP in future vehicles

2 Service-Oriented

Architectures

Diagnostics & flashing Any TCP-based

applications or protocols

Ethernet TSN

DoIP XCP Some/IP

e.g.: FTP, HTTP, SSH, SIP, car2x, cloud-based services,

electric vehicle charging,

Standard TCP/IP protocols + sockets speed-up the development of applications

requiring off/on-board reliable communications

(3)

AUTOSAR TCP/IP stacks

TCP in Automotive Ethernet Networks 3

Applications – AUTOSAR SWCs

AUTOSAR RTE

AUTOSAR TCP/IP

Socket Adaptor signals PDUs

sockets

UDP TCP

IPv4 / IPv6

Ethernet interface datagrams

ADAPTIVE AUTOSAR OS TCP/IP

Ethernet interface datagrams UDP TCP

IPv4 / IPv6

Adaptive Applications

sockets

network binding AUTOSAR ARA

Runtime for Adaptive Applications

(4)

Objectives

1. AUTOSAR TCP/IP design choices

2. Maximum achievable TCP performances with & without interfering traffic

3. Guidelines for configuring AUTOSAR TCP/IP for on-board communication

4. Impact of shapers on TCP traffic: illustration with CBS used for video

TCP performances and configuration has been studied for 40+ years, but

what about TCP – as specified by AUTOSAR – for in-vehicle communication ?

Important study in the literature: “On AUTOSAR TCP/IP Performance in In-Vehicle Network Environments”, in IEEE Communications Magazine, vol. 54, no. 12, pp. 168-173, Dec. 2016.

(5)

Techniques & toolset

– Worst-case Traversal Time (WCTT) analysis – for hard deadline constraints – Timing-accurate Simulation – for TCP throughput constraints

– Optimization algorithms for setting the parameters of all supported protocols

– RTaW-Pegase: modeling / analysis /

configuration of automotive Ethernet TSN – AUTOSAR TCP/IP stack model implemented

in RTaW-Pegase

Toolset

(6)

Case-study : Network topology

6

7 video streams (30 FPS) for vision and ADAS

1Gbit/s

↖18%

↗35% 26%↙

→30%

#Nodes 8

#Switches 2

#streams 47

+ TCP streams

Load per link (wo TCP streams)

Min: <1%, max:35%

Link data rates 100Mbit/s and 1Gbit/s (1 link)

→26%

↖16%

 TCP connection from ECU1 to ECU6

 Loads shown for interfering traffic only – not TCP streams

TCP in Automotive Ethernet Networks

(7)

Case-study : Traffic

7 Command & Control (CC)

Video streams Audio Streams

TCP streams

 32 streams, 256 to 1024 byte frames

 5ms to 80ms period and deadlines

 Hard deadline constraints

 8 streams: 128 and 256 byte frames

 1.25ms period and deadline

 deadline constraints (soft)

 3 streams (vision): 30x1400 byte frames every 33ms – deadline = 33ms

 4 streams (ADAS): 15x1000bytes frames every 33ms – deadline = 10ms

 hard and soft deadline constraints

 Bulk data = from 64K to 1MB transfers, or

 100ms periodic PDUs data, e.g. from CAN networks

Top priority

Second priority level

Lowest priority Third priority level

(8)

AUTOSAR TCP specification

8

(9)

AUTOSAR TCP design choices

– A full-fledged TCP implementation!

– Not included in the specification: selective ack (sack) and timestamp

options, recent congestion control algorithms (BIC, CUBIC, PRR, BBR)

Our view: sound design choices but

configuration is difficult because

- application specific

- subtle interactions between parameters:

e.g. send/receive windows size, TCP task

period, Nagle’s algorithm on/off, time-out

(10)

Bulk traffic: Nagle’s algorithm and delayed ack

– Improve TCP efficiency by postponing both sending of data and sending of ack → buffering on both the sending and receiving sides

64kB = first 43 segments of 1460bytes … 200ms later comes last segment  because of Nagle’s algorithm, TCP waits for the delayed ack (200ms). Solution in Autosar is to turn off “Nagle”.

200ms ?!

(11)

PDU traffic: Nagle detrimental as well

– 3 PDU streams over a TCP connection | 8, 20 and 64bytes at the lowest priority level – Maximum latency: from the time the PDU is written in the socket, until receiver reads it

0 5 10 15 20 25 30

TCP traffic alone TCP + interfering

trafic

Maximum latency (ms)

PDU maximum latencies with and without Nagle

Nagle OFF Nagle ON

For PDU’s as well, solution in

Autosar is to turn off Nagle’s

algorithm

(12)

Max. achievable performances with TCP

– throughput for bulk traffic

– latencies for PDU traffic

12

(13)

Throughput – no interfering traffic

– Experimental conditions: all mechanisms on but Nagle, event-triggered

management of TCP stacks, receive window larger than data, no packet loss

 Max. throughput is quickly reached: 96Mbps of TCP data over 100Mbps links!

 With interfering traffic (not shown), remaining available bandwidth can be fully used too

 But no exponential increase during slow-start ?!

(14)

Almost no “slow-start” phase

0 10000 20000 30000 40000 50000 60000 70000

0 2000 4000 6000 8000 10000 12000 14000 16000 18000

sequence number

time

"exponential" slow start

usable window cwnd sent seq acked seq

0 10000 20000 30000 40000 50000 60000 70000

0 1000 2000 3000 4000 5000 6000

sequence number

time

"linear" slow start

usable window queued seq sent seq acked seq cwnd

Expected: throughput growth Observed: max. throughput from start

 Reduced automotive round-trip times changes the usual behavior of TCP

 Re-examine what we can expect from TCP in the automotive context

[typical automotive round-trip times]

[Round-trip times of 2ms]

(15)

PDU latencies vs receiver reading period

– PDU TCP streams = 8, 20 and 64 bytes at the lowest priority level

– Maximum latency | Nagle off | window update sent asap after receiver reads buffer

PDU maximum latencies can be

controlled by adjusting

receiver’s reading period

0 5 10 15 20 25 30 35

5 10 15 20

Maximum latency (ms)

Receiver reading period (ms)

PDU maximum latencies vs receiver reading period TCP traffic alone TCP + interfering traffic

Maximum latency ≈ traversal time + reading period

TCPMainFunction may further delay data transfer to application

(16)

TCP configuration in a TSN network

16

(17)

Experimental setup

– Interfering streams configured so as to meet their latency constraints

– Video under CBS configured with Tight-IdleSlope algorithm = minimum Idle-

Slopes allowing to meet deadline constraints

– TCP traffic: 1MB transfers (=685 segments) between ECU1 and ECU6 every 1s

– Minimum throughput over all TCP transfers collected over long simulations:

sample of 36000 data points (12 hours of functioning)

– Receiver reads TCP buffer every 1ms

Config #2: End-to-end shaping with TSN Credit-Based Shaper (CBS)

Config #1: video streams not shaped

(18)

Shaping improves TCP throughput and reduces switch memory usage

18

CBS improves TCP throughput (up to 30%) and reduces memory requirement (up to 14%) for all parameters – larger gains with smaller TCP transfers more subject to bursts of interfering traffic

80 90 100 110 120 130 140 150 160

30 35 40 45 50 55 60 65 70

Memory used in switches (Kb)

TCP throughput (Mbps)

Throughput vs maximum memory for varying receive window sizes (rcwnd)

no CBS CBS for video

rcwnd=5MSS

rcwnd=64KB rcwnd values:

5MSS 8MSS 10MSS 11MSS 20MSS 30MSS 64KB

Larger rcwnd improves throughput but

more memory required in the switches to not

lose packets

(19)

Configuring TCP receive window size – efficiency areas

19

With or without CBS, larger receive windows improve throughput – the gain drops after a threshold that depends on how often receive buffer is read

rcwnd values:

5MSS 8MSS 10MSS 11MSS 20MSS 30MSS 64KB

80 90 100 110 120 130 140 150 160

30 35 40 45 50 55 60 65 70

Memory used in switches (Kb)

Throughput vs maximum memory for varying receive window sizes (rcwnd)

no CBS CBS for video

rcwnd=5MSS

rcwnd=64KB Increasing rcwnd size is effective:

4Mbps per additional MSS

much less effective after threshold:

0.13Mbps per MSS

(20)

In practice, larger receive windows can be detrimental!

– Memory for switch port SW2 to ECU6 set to 30Kb, packet is dropped if memory full – TCP bulk traffic |average latency | Nagle off

Receive window size should be set wrt switch memory

0 10 20 30 40 50 60 70

20 21 22 23 24 25 26

Receive window size in MSS

TCP throughput vs receive window size

without CBS with CBS

Larger receive windows means more “in-flight” data.

Packet losses in switches lead to retransmission after

time-out (1s) and drop in throughput!

(21)

Takeways

1

2

3

AUTOSAR TCP is able to use all of the available bandwidth with minimal latencies – if properly configured and enough memory

 TCP for soft real-time only as one can just obtain statistical guarantees (i.e., no worst-case analysis)

 The use of TSN shapers at higher priority levels improves TCP performance and reduces overall memory requirement

AUTOSAR specifies a full-fledged TCP protocol Need to re-examine what we know about TCP in the

automotive context

AUTOSAR TCP configuration choices make a huge difference, parameters cannot be set in isolation

 E.g. best choices for receive window size & polling period depend on switch memory size

(22)

22

Thank you for your attention!

Questions? Feedback? contact us at jorn.migge@realtimeatwork.com

nicolas.navet@uni.lu

(23)

References

23

(24)

References

1. S. Han and H. Kim, “On AUTOSAR TCP/IP Performance in In-Vehicle Network Environments”, in IEEE Communications Magazine, vol. 54, no. 12, pp. 168-173, Dec. 2016.

2. AUTOSAR, “Specification of TCP/IP Stack”, release 4.3.1., 2017.

3. J. Migge, M. Boyer, N. Navet, J. Villanueva, “Insights on the performance and configuration of AVB and TSN in automotive networks“, Proc. Embedded Real-Time Software and Systems (ERTS 2018), Toulouse, France, January, 2018. Available at http://www.realtimeatwork.com/wp-

content/uploads/Ethernet-QoS_RTaW.pdf.