Performance hit with IP-tunnels

Performance hit with IP-tunnels

[Kristian Evensen]

Hello,

I am currently comparing different IP-tunneling protocols/implementations,
and have stumbled upon something I am not able to explain. Regardless of
which tunneling technology I use, the latency increases with a couple of 10s
of ms and I see a significant degradation of throughput (compared to when no
tunnels are used). The only exception is IP-in-IP, where I get similar
performance with and without tunnels, but it does unfortunately not work in
my scenario.

First, I thought this was caused by the different tunneling software, but
after measuring the processing time of the applications (xl2tp and
pptp-client) and when the packets are seen by the different iptables chains
(using LOG), these delays seem to be acceptable. However, one delay sticks
out. After the packet has been decapsulated and fed to PPP, it takes a
"long" time before it is seen again. My question is, can PPP be the cause of
the higher latency and lower throughput? 

Similar observations are made at both ends of the tunnel.

Thanks in advance for any help,
Kristian

Performance hit with IP-tunnels

[Kristian Evensen]

Hello,

I am currently comparing different IP-tunneling protocols/implementations,
and have stumbled upon something I am not able to explain. Regardless of
which tunneling technology I use, the latency increases with a couple of 10s
of ms and I see a significant degradation of throughput (compared to when no
tunnels are used). The only exception is IP-in-IP, where I get similar
performance with and without tunnels, but it does unfortunately not work in
my scenario.

First, I thought this was caused by the different tunneling software, but
after measuring the processing time of the applications (xl2tp and
pptp-client) and when the packets are seen by the different iptables chains
(using LOG), these delays seem to be acceptable. However, one delay sticks
out. After the packet has been decapsulated and fed to PPP, it takes a
"long" time before it is seen again. My question is, can PPP be the cause of
the higher latency and lower throughput? 

Similar observations are made at both ends of the tunnel.

Thanks in advance for any help,
Kristian

Re: Performance hit with IP-tunnels

[Eric Dumazet]

Le lundi 22 mars 2010 à 11:17 +0100, Kristian Evensen a écrit :
> Hello,
> 
> I am currently comparing different IP-tunneling protocols/implementations,
> and have stumbled upon something I am not able to explain. Regardless of
> which tunneling technology I use, the latency increases with a couple of 10s
> of ms and I see a significant degradation of throughput (compared to when no
> tunnels are used). The only exception is IP-in-IP, where I get similar
> performance with and without tunnels, but it does unfortunately not work in
> my scenario.
> 
> First, I thought this was caused by the different tunneling software, but
> after measuring the processing time of the applications (xl2tp and
> pptp-client) and when the packets are seen by the different iptables chains
> (using LOG), these delays seem to be acceptable. However, one delay sticks
> out. After the packet has been decapsulated and fed to PPP, it takes a
> "long" time before it is seen again. My question is, can PPP be the cause of
> the higher latency and lower throughput? 
> 
> Similar observations are made at both ends of the tunnel.

A soon as a round trip on a user process is requested to handle a
packet, you can have delay because of scheduling constraints.

You could try latencytop and check if something strange raises, 10 ms
seems excessive.

IP-TIP tunnels dont use a user space program, they are immune to
scheduler latencies.

Re: Performance hit with IP-tunnels

[Eric Dumazet]

Le lundi 22 mars 2010 à 11:17 +0100, Kristian Evensen a écrit :
> Hello,
> 
> I am currently comparing different IP-tunneling protocols/implementations,
> and have stumbled upon something I am not able to explain. Regardless of
> which tunneling technology I use, the latency increases with a couple of 10s
> of ms and I see a significant degradation of throughput (compared to when no
> tunnels are used). The only exception is IP-in-IP, where I get similar
> performance with and without tunnels, but it does unfortunately not work in
> my scenario.
> 
> First, I thought this was caused by the different tunneling software, but
> after measuring the processing time of the applications (xl2tp and
> pptp-client) and when the packets are seen by the different iptables chains
> (using LOG), these delays seem to be acceptable. However, one delay sticks
> out. After the packet has been decapsulated and fed to PPP, it takes a
> "long" time before it is seen again. My question is, can PPP be the cause of
> the higher latency and lower throughput? 
> 
> Similar observations are made at both ends of the tunnel.

A soon as a round trip on a user process is requested to handle a
packet, you can have delay because of scheduling constraints.

You could try latencytop and check if something strange raises, 10 ms
seems excessive.

IP-TIP tunnels dont use a user space program, they are immune to
scheduler latencies.