Using MultiPath TCP to enhance home networks

Over the last few months I’ve been playing with MultiPath TCP and in this post I will show how I use it to leverage my humble True ADSL line at home.

For performance and security reasons, I tunnel all my traffic thru a VPN. This is not necessarily to circumvent censorship, but to circumvent the evil transparent proxies my ISP puts in middle. The total bandwidth available is ~10 mbps down / ~1 mbps up.

Introduction to MultiPath TCP

MultiPath TCP is an interesting effort to use multiple interfaces/networks for any single TCP connection. A Linux kernel implementation is being developed at Its main use cases are for mobile (transition between Wi-Fi and 3G) and datacenters. I exploit it to get better Internet browsing experience.

Old way

Simple SSH Tunnel

  1. SSH tunnel to EC2 instance in Singapore. Browser configured to use this tunnel as proxy
  2. SSH tunnel to EC2 instance in us-east (for accessing geo-blocked services)

Main drawback : If my ISP has issues talking to AWS, then I’m totally screwed. This happened a month or so ago where most links coming into True was severely limited, however link from Digital Ocean to True was healthy. I had to manually change my tunnels to a $5 Digital Ocean instance.

New way

MPTCP Tunnel

Note: This is a constantly evolving setup as I find new things to play with.

Infrastructure involved :-

All TCP Traffic is intercepted by the APU using iptables, diverted to redsocks, which sends it to the shadowsocks client, which sends it to the shadowsocks server running in EC2 Singapore. This socks connection has several ways to communicate with the EC2 instance.

APU <--> True ADSL Directly <--> EC2  
APU <--> True ADSL Directly over OpenVPN/UDP <--> EC2  
APU <--> True ADSL <--> via CAT VPS over OpenVPN/UDP <--> EC2  
APU <--> True ADSL <--> via DO Singapore over OpenVPN/UDP <--> EC2  
APU <--> Dtac 3G Directly <--> EC2 (Optional/ondemand)

Now I have 5 possible paths. MPTCP kernel creates a TCP connection over each available path and bonds them together and exposes it as a single TCP connection to the application. Packets are sent over paths that currently have the lowest delay. Now my available bandwidth is not impacted by congestion over some of these paths. All paths need to be congested for me to have a bad day… Also some path might have good uplink, some might have good downlink, with MPTCP you mix the best of both…

Example bmon stats when downloading a large file (I removed irreverent interfaces.)

  #   Interface                RX Rate         RX #     TX Rate         TX #
xxx (source: local)
  0   tun1                     621.28KiB        628      38.82KiB        636
  3   tun3                     200.22KiB        198       9.42KiB        149
  5   ppp0                       1.07MiB       1018     119.42KiB        980
  9   tun0                      90.06KiB         90       5.94KiB         97



Jumpbox is pretty basic setup. It’s role is to provide additional gateways which MPTCP uses to build additional paths.

OpenVPN server configured normally. Set to not redirect default gateway. In my current setup I need to ensure that the server assigns the same IP to my client. This is not really that crucial, but it keeps things simple. Its important to configure each jumpbox to use a different IP range.

net.ipv4.ip_forward needs to be set to 1 to allow forwarding. In fact almost all boxes in the setup need this.

iptables rules needed :-

iptables -t nat -A POSTROUTING -o eth0 -j MASQUERADE
iptables -t filter -A FORWARD -i tun0 -j ACCEPT
iptables -t filter -A FORWARD -o tun0 -j ACCEPT

Replace the tun0 and eth0 to suit your environment.


A destination server is remote end of our socks tunnel. It’s job is to service the socks connections patching them to the real destination.

This needs to run a MultiPath TCP kernel. On EC2 it is pretty simple. Launch an Ubuntu 14.04 instance with a pv-grub AKI. Then follow the apt-repository installation method. And ensure the grub loads the MPTCP kernel as its first choice.

Next we also need shadowsocks server running. RTFM its pretty simple. Before using shadowsocks I was using a simple ssh -D tunnel, but I found it to be inefficient. Often times one large transfer would make all other TCP streams stuck. Perhaps this has something to do with the fact that with SSH everything is happening over a single TCP stream whereas shadowsocks makes a new socks connection dedicated to each TCP connection.


The gateway is the most complicated component. Running stock Debian wheezy with MPTCP kernel installed via their apt repository. A lot of services run here. I will not elaborate on some of them.

dbcpd - Assign LAN users with IP

bind - For DNS recursion. Since we tunnel most traffic to Singapore, I also set bind to send DNS queries thru OpenVPN.

iptables - I use iptables to do the NAT. NAT all UDP packets to OpenVPN. Send all outgoing TCP connections to redsocks.

# Generated by iptables-save v1.4.14 on Sat Nov 22 00:37:10 2014
:PREROUTING ACCEPT [378881:57485495]
:INPUT ACCEPT [210208:17266788]
:OUTPUT ACCEPT [4099955:310913862]
:POSTROUTING ACCEPT [3239510:252587265]
:REDSOCKS - [0:0]
- -A PREROUTING -i br0 -p tcp -j REDSOCKS
- -A REDSOCKS -d a.b.c.d/32 -j RETURN
- -A REDSOCKS -d e.f.g.h/32 -j RETURN
- -A REDSOCKS -d i.j.k.l/32 -j RETURN
- -A REDSOCKS -d m.n.o.p/32 -j RETURN
- -A REDSOCKS -d q.r.s.t/32 -j RETURN
- -A REDSOCKS -s -p tcp -j REDIRECT --to-ports 12345
# Completed on Sat Nov 22 00:37:10 2014
# Generated by iptables-save v1.4.14 on Sat Nov 22 00:37:10 2014
:INPUT ACCEPT [115657469:73738905421]
:FORWARD ACCEPT [64078:47442189]
:OUTPUT ACCEPT [122121802:63701527314]
- -A FORWARD -i eth1 -j ACCEPT
- -A FORWARD -o eth1 -j ACCEPT
- -A FORWARD -i br0 -j ACCEPT
- -A FORWARD -o br0 -j ACCEPT
- -A FORWARD -i eth0 -j ACCEPT
- -A FORWARD -o eth0 -j ACCEPT
# Completed on Sat Nov 22 00:37:10 2014
# Generated by iptables-save v1.4.14 on Sat Nov 22 00:37:10 2014
:PREROUTING ACCEPT [118734314:74507536093]
:INPUT ACCEPT [115635709:73734306355]
:FORWARD ACCEPT [3100331:759352048]
:OUTPUT ACCEPT [122104929:63698976437]
:POSTROUTING ACCEPT [125198421:64456746251]
- -A PREROUTING ! -d -i br0 -j MARK --set-xmark 0x1/0xffffffff
- -A PREROUTING -d -i br0 -j MARK --set-xmark 0x3/0xffffffff
- -A PREROUTING -d -i br0 -j MARK --set-xmark 0x2/0xffffffff
# Completed on Sat Nov 22 00:37:10 2014

Note: a.b.c.d, e.f.g.h, i.j.k.l, m.n.o.p and q.r.s.t are public internet ips that I don’t want redsocks to intercept.


Also, each interface is maintains its own routing tables using if-up scripts. For example this is what gets executed when one of the tunnels comes alive.

ip rule add from table 2 || true
ip route add dev tun0 scope link table 2 || true
ip route add default via dev tun0 table 2 || true
ip rule add fwmark 3 table 2 || true

The fwmark is added so if in future I want to pipe different traffic to go thru this interface I can set the corresponding iptables rule.

All local services are scoped to listen only on local interfaces to avoid random people connecting to local services.

redsocks - Accepts intercepted connections and pipes it off to shadowsocks client

shadowsocks client - sends all TCP connections to shadowsocks servers running in EC2 Singapore and US. By default intercepted traffic is sent to Singapore, however any application on any computer in the network could be set to explicitly use any of the available proxies.

wvdial - To dial the ADSL connection which is itself behind a Carrier-grade NAT. Sometimes the connection stops working while pppd things its still connected. Am ugly CRON script to test the network and flip it if needed.


IP=`/sbin/ip route | grep -v default | grep ppp0 | cut -d " " -f 1`
COUNT=`echo $IP | wc -l`

echo $IP $COUNT
if [ $COUNT -eq 1 ] 
  ping -c 10 $IP > /dev/null
  if [ $? -eq 0 ]
    echo "ppp0 is up"
    echo "ppp0 is down"
    kill -SIGHUP `pgrep pppd`
    beep -l 25
  echo "ppp0 not found?!?!?!?1"
The script above tries to ping the default gateway of the ppp0 interface to find out if it is really up. SIGHUP signals pppd to handup and redial. I don’t bother maintaining the pid of pppd because currently I use ppp0 exclusively for the ADSL modem.

Missing parts

There are some issues I am having that I need to sort out work-around for.

Future path enhancements

More paths can be added to get better throughput.


MPTCP is a fantastic piece of technology. Hat-tip to everyone who contributed to it. The PC Engines ALU box is also awesome. Decent x86_64 box consuming only about 6 to 12W power.

In the future I will do a walk-thru type post on how to setup a Raspberry Pi as a one-arm gateway doing a subset of what I described above. The most challenging part is getting a MPTCP enabled kernel on the pi, which requires kernel patching and compiling. The throughput will likely be very limited because MPTCP has a higher CPU overhead than regular TCP.

Tags: mptcp tcp network