Skip to contentSkip to conversation

disabling nagle

DISCUSSION Dave Slicer 4 participants 5 messages Feb 4, 2003 – Feb 6, 2003

lore
Share & tools
lore
#0 lore raw
Quoted 4 lines · Fiona Sou-Yee Wong

I have kernel version 2.4.18 and I was looking for a patch to have the option to disable NAGLE's algorithm.
Is there a patch available for kernels 2.4 and greater and if not, what other options do I have?

Others already answered this specific question, but I wonder how hard it would be to create a patch to disable TCP's timeout and retransmit mechanisms on a given interface? This would allow those of us who have no alternative other than PPP over ssh for VPN to greatly improve performance. Over a well behaved connection this works acceptably, but given any delays or packet loss it is essentially unusable. I know the real answer is using something other than TCP as the transport layer for the tunnel (IPSEC, IP over IP, UDP, etc.) but that isn't always possible. So I'd like a way to treat the ppp interface the VPN tunnel creates as a completely reliable transport for which normal TCP/IP retransmits and timeouts don't apply. It'd just bullheadedly go along transmitting data and assuming it was received -- the underlying TCP transport can take care of making the link reliable.

Is this even remotely reasonable? If it would cause performance degradation it'd have to be a config option or never make the kernel at all (Linus may never accept it regardless I suppose) But ignoring that for a moment, is it just too hairy to contemplate? I've done a few patches here and there for Linux in the past, but nothing like this (and nothing involving networking) so it is far beyond my capability. But if something was cooked up that works well enough I'd be willing to try polishing it and porting between kernel versions where necessary.

But I'd take any suggestions for alterations in /proc/sys/net/ipv4/* that might help the current state of things.

_________________________________________________________________
Help STOP SPAM with the new MSN 8 and get 2 months FREE* http://join.msn.com/?page=features/junkmail

#1 lore raw
Quoted 6 lines · Dave Slicer
Quoted 4 lines · Fiona Sou-Yee Wong

I have kernel version 2.4.18 and I was looking for a patch to have the option to disable NAGLE's algorithm.
Is there a patch available for kernels 2.4 and greater and if not, what other options do I have?

Quoted 14 lines

Others already answered this specific question, but I wonder how hard it would be to create a patch to disable TCP's timeout and retransmit mechanisms on a given interface? This would allow those of us who have no alternative other than PPP over ssh for VPN to greatly improve performance. Over a well behaved connection this works acceptably, but given any delays or packet loss it is essentially unusable. I know the real answer is using something other than TCP as the transport layer for the tunnel (IPSEC, IP over IP, UDP, etc.) but that isn't always possible. So I'd like a way to treat the ppp interface the VPN tunnel creates as a completely reliable transport for which normal TCP/IP retransmits and timeouts don't apply. It'd just bullheadedly go along transmitting data and assuming it was received -- the underlying TCP transport can take care of making the link reliable.

I want this too ;) For one, it would be a perfect example of using good existing tools to achieve the goal instead of inventing something big and new. Also it does not reduce MTU unlike packet-encapsulation tunnels.

Now it's an imperfect example due to noted TCP over TCP performance problem ('internal meltdown').

Quoted 10 lines

Is this even remotely reasonable? If it would cause performance degradation it'd have to be a config option or never make the kernel at all (Linus may never accept it regardless I suppose) But ignoring that for a moment, is it just too hairy to contemplate? I've done a few patches here and there for Linux in the past, but nothing like this (and nothing involving networking) so it is far beyond my capability. But if something was cooked up that works well enough I'd be willing to try polishing it and porting between kernel versions where necessary.

Quoted 2 lines

But I'd take any suggestions for alterations in /proc/sys/net/ipv4/* that might help the current state of things.

I'm looking there for the first time ever, but it seems you can twiddle TCP parameters in /proc/sys/net/ipv4/tcp_* (OTOH I don't see retransmit timeout controls there... maybe they have another name?)

In order to make them per-iface one needs to have an ability to override them in /proc/sys/net/ipv4/conf/ethX.
Seems like this is not implemented.
--
vda

#2 lore raw
Quoted 16 lines · Denis Vlasenko
Quoted 6 lines · Dave Slicer

Others already answered this specific question, but I wonder how hard it would be to create a patch to disable TCP's timeout and retransmit mechanisms on a given interface? This would allow those of us who have no alternative other than PPP over ssh for VPN to greatly improve performance. Over a well behaved connection this works acceptably, but given any delays or packet loss it is essentially

Quoted 8 lines

unusable. I know the real answer is using something other than TCP as the transport layer for the tunnel (IPSEC, IP over IP, UDP, etc.) but that isn't always possible. So I'd like a way to treat the ppp interface the VPN tunnel creates as a completely reliable transport for which normal TCP/IP retransmits and timeouts don't apply. It'd just bullheadedly go along transmitting data and assuming it was received -- the underlying TCP transport can take care of making the link reliable.

Quoted 7 lines

I want this too ;) For one, it would be a perfect example of using good existing tools to achieve the goal instead of inventing something big and new. Also it does not reduce MTU unlike packet-encapsulation tunnels.

Now it's an imperfect example due to noted TCP over TCP performance problem ('internal meltdown').

I doubt a hack like disabling RTO would make it into the kernel. However, try enabling F-RTO at both ends (echo 1 > /proc/sys/net/ipv4/tcp_frto). This should improve things quite a bit.

You need at least linux 2.4.21-pre3, or linux 2.5.x.

MikaL

#3 lore raw
Quoted 15 lines
Quoted 15 lines
Quoted 9 lines

loss it is essentially unusable. I know the real answer is using something other than TCP as the transport layer for the tunnel (IPSEC, IP over IP, UDP, etc.) but that isn't always possible. So I'd like a way to treat the ppp interface the VPN tunnel creates as a completely reliable transport for which normal TCP/IP retransmits and timeouts don't apply. It'd just bullheadedly go along transmitting data and assuming it was received -- the underlying TCP transport can take care of making the link reliable.

I want this too ;) For one, it would be a perfect example of using good existing tools to achieve the goal instead of inventing something big and new. Also it does not reduce MTU unlike packet-encapsulation tunnels.

Quoted 8 lines
Quoted 2 lines

Now it's an imperfect example due to noted TCP over TCP performance problem ('internal meltdown').

I doubt a hack like disabling RTO would make it into the kernel. However, try enabling F-RTO at both ends (echo 1 > /proc/sys/net/ipv4/tcp_frto). This should improve things quite a bit.

You need at least linux 2.4.21-pre3, or linux 2.5.x.

Wow, thanks! I'll look into it! (Found the draft. Google is cool ;)

What is really needed is raising RTO to large fixed value so that TCP connection times out before RTs are triggered. And it have to be done only on pppd-over-ssh iface. Am I right that currently kernel TCP options are global, not per-if?
--
vda

#4 lore raw
In reply to Dave Slicer
Quoted 9 lines

would be to create a patch to disable TCP's timeout and retransmit mechanisms on a given interface? This would allow those of us who have no alternative other than PPP over ssh for VPN to greatly improve performance. Over a well behaved connection this works acceptably, but given any delays or packet loss it is essentially unusable. I know the real answer is using something other than TCP as the transport layer for the tunnel (IPSEC, IP over IP, UDP, etc.) but that isn't always possible. So I'd like a way to treat the ppp interface the VPN tunnel creates as a completely reliable transport for which normal TCP/IP retransmits and timeouts don't apply.

As I see it, this wouldn't help when the TCP retransmits are originated by a machine other than the VPN routers, because they don't know about the new reliability characteristics of their transport medium. So in any network with more than two machines its usefulness is rather limited. (Or am I missing something?)

Olaf