Many applications are network I/O bound, including common database-based applications and service-based architectures. But operating systems and applications are often untuned to deliver high performance. This session uncovers hidden issues that lead to low network performance, and shows you how to overcome them to obtain the best network performance possible.

1. © 2015, Amazon Web Services, Inc. or its Affiliates. All rights reserved.
Kevin Miller, Sr. Manager, EC2 Networking
October 2015
NET404
Making Every Packet Count

2. What to Expect from this Session
Tuning TCP
on Linux
TCP Performance Application

3. What to Expect from this Session
Application
Watch us
increase network
performance
137%

5. TCP

6. TCP
• Transmission Control Protocol
• Underlies SSH, HTTP, *SQL, SMTP
• Stream delivery, flow control

7. TCP
Jack Jill

8. Jack Jill

9. Limiting in-flight data
Jack Jill
Receive
Window
Receive
Window
Congestion
Window
Congestion
Window
Round trip time

10. Bandwidth delay product
Jack Jill
2 ms round-trip time

11. Bandwidth delay product
Jack Jill
100 ms round-trip time

12. Receive window
Receiver controlled, signaled to sender

13. Congestion window
Jack Jill
Receive
Window
Receive
Window
Congestion
Window
Congestion
Window
Round trip time

14. Congestion window
• Sender controlled
• Window is managed by the congestion control algorithm
• Inputs – varies by algorithm


15. Initial congestion window
$ ip route list
default via 10.16.16.1 dev eth0
10.16.16.0/24 dev eth0 proto kernel scope link
169.254.169.254 dev eth0 scope link
1448 1448 1448 = 4344 bytes

16. Initial congestion window
# ip route change 10.16.16.0/24 dev eth0
proto kernel scope link initcwnd 16
$ ip route list
default via 10.16.16.1 dev eth0
10.16.16.0/24 dev eth0 proto kernel scope link
initcwnd 16
169.254.169.254 dev eth0 scope link
1448 1448 1448 1448[ + 12 ] = 23168 bytes

17. 0
20
40
60
80
100
0% 2% 4% 6% 8% 10%
Loss Rate
Impact of loss on TCP throughput

18. Loss is visible as TCP retransmissions
$ netstat -s | grep retransmit
58496 segments retransmitted
52788 fast retransmits
135 forward retransmits
3659 retransmits in slow start
392 SACK retransmits failed

19. Socket level diagnostic
$ ss -ite
State Recv-Q Send-Q Local Address:Port Peer Address:Port
ESTAB 0 3829960 10.16.16.18:https 10.16.16.75:52008
timer:(on,012ms,0) uid:498 ino:7116021 sk:0001c286 <->
ts sack cubic wscale:7,7 rto:204 rtt:1.423/0.14 ato:40
mss:1448 cwnd:138 ssthresh:80 send 1123.4Mbps unacked:138
retrans:0/11737 rcv_space:26847
TCP State

20. Socket level diagnostic
$ ss -ite
State Recv-Q Send-Q Local Address:Port Peer Address:Port
ESTAB 0 3829960 10.16.16.18:https 10.16.16.75:52008
timer:(on,012ms,0) uid:498 ino:7116021 sk:0001c286 <->
ts sack cubic wscale:7,7 rto:204 rtt:1.423/0.14 ato:40
mss:1448 cwnd:138 ssthresh:80 send 1123.4Mbps unacked:138
retrans:0/11737 rcv_space:26847
Bytes queued for
transmission

21. Socket level diagnostic
$ ss -ite
State Recv-Q Send-Q Local Address:Port Peer Address:Port
ESTAB 0 3829960 10.16.16.18:https 10.16.16.75:52008
timer:(on,012ms,0) uid:498 ino:7116021 sk:0001c286 <->
ts sack cubic wscale:7,7 rto:204 rtt:1.423/0.14 ato:40
mss:1448 cwnd:138 ssthresh:80 send 1123.4Mbps unacked:138
retrans:0/11737 rcv_space:26847
Congestion
control algorithm

22. Socket level diagnostic
$ ss -ite
State Recv-Q Send-Q Local Address:Port Peer Address:Port
ESTAB 0 3829960 10.16.16.18:https 10.16.16.75:52008
timer:(on,012ms,0) uid:498 ino:7116021 sk:0001c286 <->
ts sack cubic wscale:7,7 rto:204 rtt:1.423/0.14 ato:40
mss:1448 cwnd:138 ssthresh:80 send 1123.4Mbps unacked:138
retrans:0/11737 rcv_space:26847
Retransmission
timeout

23. Socket level diagnostic
$ ss -ite
State Recv-Q Send-Q Local Address:Port Peer Address:Port
ESTAB 0 3829960 10.16.16.18:https 10.16.16.75:52008
timer:(on,012ms,0) uid:498 ino:7116021 sk:0001c286 <->
ts sack cubic wscale:7,7 rto:204 rtt:1.423/0.14 ato:40
mss:1448 cwnd:138 ssthresh:80 send 1123.4Mbps unacked:138
retrans:0/11737 rcv_space:26847
Congestion
window

24. Socket level diagnostic
$ ss -ite
State Recv-Q Send-Q Local Address:Port Peer Address:Port
ESTAB 0 3829960 10.16.16.18:https 10.16.16.75:52008
timer:(on,012ms,0) uid:498 ino:7116021 sk:0001c286 <->
ts sack cubic wscale:7,7 rto:204 rtt:1.423/0.14 ato:40
mss:1448 cwnd:138 ssthresh:80 send 1123.4Mbps unacked:138
retrans:0/11737 rcv_space:26847
Retransmissions

25. Monitoring retransmissions in real time
• Observable using Linux kernel tracing
# tcpretrans
TIME PID LADDR:LPORT -- RADDR:RPORT STATE
03:31:07 106588 10.16.16.18:443 R> 10.16.16.75:52291 ESTABLISHED
https://github.com/brendangregg/perf-tools/

26. Congestion control algorithm
Jack Jill

27. Congestion control algorithms in Linux
• New Reno: Pre-2.6.8
• BIC: 2.6.8 – 2.6.18
• CUBIC: 2.6.19+
• Pluggable architecture
• Other algorithms often available
• Vegas, Illinois, Westwood, Highspeed, Scalable

28. Tuning congestion control algorithm
$ sysctl net.ipv4.tcp_available_congestion_control
net.ipv4.tcp_available_congestion_control = cubic reno
$ find /lib/modules -name tcp_*
[…]
# modprobe tcp_illinois
$ sysctl net.ipv4.tcp_available_congestion_control
net.ipv4.tcp_available_congestion_control = cubic reno illinois

29. Tuning congestion control algorithm
# sysctl net.ipv4.tcp_congestion_control=illinois
net.ipv4.tcp_congestion_control = illinois
# echo “net.ipv4.tcp_congestion_control = illinois” >
/etc/sysctl.d/01-tcp.conf
[Restart network processes]

30. Retransmission timer
• Input to when the congestion control
algorithm considers a packet lost
• Too low: spurious retransmission; congestion control can
over-react and be slow to re-open the congestion
window
• Too high: increased latency while algorithm determines a
packet is lost and retransmits

31. Tuning retransmission timer minimum
• Default minimum: 200ms
# ip route list
default via 10.16.16.1 dev eth0
10.16.16.0/24 dev eth0 proto kernel scope link
169.254.169.254 dev eth0 scope link
Route to other
instances in
our subnet
(same AZ)

32. Tuning retransmission timer minimum
# ip route list
default via 10.16.16.1 dev eth0
10.16.16.0/24 dev eth0 proto kernel scope link
169.254.169.254 dev eth0 scope link
# ip route change 10.16.16.0/24 dev eth0 proto kernel
scope link rto_min 10ms
# ip route list
default via 10.16.16.1 dev eth0
10.16.16.0/24 dev eth0 proto kernel scope link rto_min
lock 10ms
169.254.169.254 dev eth0 scope link

33. Queueing along the network path
Jack Jill

34. Queueing along the network path
• Intermediate routers along a path have
interface buffers
• High load leads to more packets in buffer
• Latency increases due to queue time
• Can trigger retransmission timeouts

35. Active queue management
$ tc qdisc list
qdisc mq 0: dev eth0 root
qdisc pfifo_fast 0: dev eth0 parent :1 bands 3 […]
qdisc pfifo_fast 0: dev eth0 parent :2 bands 3 […]
# tc qdisc add dev eth0 root fq_codel
qdisc fq_codel 8006: dev eth0 root refcnt 9 limit 10240p
flows 1024 quantum 9015 target 5.0ms interval 100.0ms ecn
http://www.bufferbloat.net/projects/codel/wiki

36. Amazon EC2 enhanced networking
Jack Jill

37. Amazon EC2 enhanced networking
• Higher I/O (packets per second) performance
• Lower CPU utilization
• Lower inter-instance latency
• Low network jitter
• Instance families: M4, C4, C3, R3, I2, D2 (w/ HVM)
• Drivers built into Windows, Amazon Linux AMIs
• Questions? re:Invent 2014 – SDD419

38. Maximum transmission unit
3.47% overhead vs. 0.58% overhead
Improvement seen among instances in your VPC
1448B
Payload
8949B Payload

39. Tuning maximum transmission unit
# ip link list
2: eth0: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 9001 qdisc
mq state UP mode DEFAULT group default qlen 1000
link/ether 06:f1:b7:e1:3b:e7
# ip route list
default via 10.16.16.1 dev eth0
10.16.16.0/24 dev eth0 proto kernel scope link
169.254.169.254 dev eth0 scope link

40. Tuning maximum transmission unit
# ip route change default via 10.16.16.1 dev eth0 mtu 1500
# ip route list
default via 10.16.16.1 dev eth0 mtu 1500
10.16.16.0/24 dev eth0 proto kernel scope link
169.254.169.254 dev eth0 scope link

41. Applying our new knowledge

42. Test setup
• m4.10xlarge instances – Jack and Jill
• Amazon Linux 2015.09 (Kernel 4.1.7-15.23.amzn1)
• Web Server: nginx 1.8.0
• Client: ApacheBench 2.3
• TLSv1,ECDHE-RSA-AES256-SHA,2048,256
• Transferring uncompressible data (random bits)
• Origin data stored in tmpfs (RAM based; no server disk I/O)
• Data discarded once retrieved (no client disk I/O)

43. Example Apache Bench output
[ … ]
Concurrency Level: 100
Time taken for tests: 59.404 seconds
Complete requests: 10000
Failed requests: 0
Write errors: 0
Total transferred: 104900000 bytes
HTML transferred: 102400000 bytes
Requests per second: 168.34 [#/sec] (mean)
Time per request: 594.038 [ms] (mean)
Time per request: 5.940 [ms] (mean, across all
concurrent requests)
Transfer rate: 1724.49 [Kbytes/sec] received
[ … ]

44. Application 1
HTTPS with intermediate network loss
Jack Jill
0.2%
loss

45. Test setup
• 1 test server instance, 1 test client instance
• 80ms RTT
• 160 parallel clients retrieving a 100 MB object 5 times
$ ab -n 100 -c 20 https://server/100m [* 8]
• Simulated packet loss
# tc qdisc add dev eth0 root netem loss 0.2%
Goal: Minimize throughput impact with 0.2% loss

46. Results – application 1
Test Bandwidth Mean Time
All defaults – no loss 4163 Mbps 27.9s
All defaults – 0.2% simulated loss 1469 Mbps 71.8s
Increased initial congestion window w/ loss 1328 Mbps 80.6s
Doubled server-side TCP buffers w/ loss 1366 Mbps 78.6s
Illinois congestion control algorithm w/ loss 3486 Mbps 28.2s
137% increase
in performance!

47. Application 2
Bulk data transfer; high RTT path
Jack Jill

48. Test setup
• 1 test server instance, 1 test client instance
• 80 ms RTT
• 8 parallel clients retrieving a 1 GB object 2 times
$ ab -n 2 -c 1 https://server/1g [* 8]
Goal: Maximize the throughput / minimize transfer time

49. Results – application 2
Test Bandwidth Mean Time
All defaults 2164 Mbps 30.4s
Doubled TCP buffers on server end 1780 Mbps 37.4s
Doubled TCP buffers on client end 2462 Mbps 27.6s
Active queue management on server 2249 Mbps 29.3s
Client buffers + AQM 2730 Mbps 24.5s
Illinois CC + client buffers + AQM 2847 Mbps 23.0s
Illinois CC + server & client buffers + AQM 2865 Mbps 23.5s
32% increase in
performance!

50. Application 3
Bulk data transfer; low RTT path
Jack Jill

51. Test setup
• 1 test server instance, 1 test client instance
• 1.2 ms RTT
• 8 parallel clients retrieving a 10GB object 2 times
$ ab -n 2 -c 1 https://server/100m [* 8]
• Start at Internet default MTU, then increase
Goal: Maximize the throughput / minimize transfer time

52. Results
Test Bandwidth Mean Time
All defaults + 1500B MTU 8866 Mbps 74.0s
9001B MTU 9316 Mbps 70.4s
Active Queue Management (+MTU) 9316 Mbps 70.4s
5% increase

53. Application 4
High transaction rate HTTP service
Jack Jill

54. Test setup
• 1 test server instance, 1 test client instance
• 80 ms RTT
• HTTP, not HTTPS
• 6400 parallel clients retrieving a 10k object 100 times
$ ab -n 20000 -c 200 http://server/10k [* 32]
Goal: Minimize latency

55. Results – application 4
Test Bandwidth Mean Time
All defaults 2580 Mbps 195.3ms
Initial congestion window – 16 packets 2691 Mbps 189.2ms
Illinois CC + initial congestion window 2649 Mbps 186.2ms
4.6% decrease

56. Take-aways

57. Take-aways
• The network doesn’t have to be a black box – Linux tools
can be used to interrogate and understand
• Simple tweaks to settings can dramatically increase
performance – test, measure, change
• Understand what your application needs from the
network, and tune accordingly

58. Remember to complete
your evaluations!

59. Thank you!