is tcp_nodelay not an option ? anupam On Mon, May 30, 2011 at 9:21 AM, C K Kashyap <ckkashyap@gmail.com> wrote:

On Mon, May 30, 2011 at 9:18 AM, Anupam Kapoor <anupam.kapoor@gmail.com> wrote:

...

afaik, it disables tcp-prequeue nothing to do with nagle anyways. also, doing this might have marginal benefits if anything at all due to most overhead being specific to context switching than anything else.

Thanks Anupam ... can you pleast tell me how I could disable nagle's algo? At a system level. I've seen a lot of info on socket param on the net. Regards, Kashyap

-- In the beginning was the lambda, and the lambda was with Emacs, and Emacs was the lambda.

On Mon, May 30, 2011 at 12:47 PM, C K Kashyap <ckkashyap@gmail.com> wrote:

On Mon, May 30, 2011 at 9:53 AM, Anupam Kapoor <anupam.kapoor@gmail.com> wrote:

...

is tcp_nodelay not an option ?

Is this a socket option or is there a system wide setting for tcp_nodelay?

Yes, TCP_NODELAY is a socket option: Going back to the kernel source: include/linux/tcp.h: #define TCP_NODELAY 1 /* Turn off Nagle's algorithm. */ and this: u8 nonagle : 4,/* Disable Nagle algorithm? */ and looking into the kernel source code: ./fs/ocfs2/cluster/tcp.c: ret = sock->ops->setsockopt(sock, SOL_TCP, TCP_NODELAY, ./net/rds/tcp.c: sock->ops->setsockopt(sock, SOL_TCP, TCP_NODELAY, (char __user *)&val, the above are just two of the examples of how to set TCP_NODELAY in the kernel. At the userspace level: http://www.linuxquestions.org/questions/other-*nix-55/how-do-i-disable-the-n... should be the correct way, except that the poster is attempting to do it in a kernel that does not support it. As documented in http://linux.die.net/man/7/tcp, the supported kernel is 2.5.71 and onwards. -- Regards, Peter Teoh

On Mon, May 30, 2011 at 4:07 PM, C K Kashyap <ckkashyap@gmail.com> wrote:

...

Yes, TCP_NODELAY is a socket option:

Going back to the kernel source:

include/linux/tcp.h:

#define TCP_NODELAY             1       /* Turn off Nagle's algorithm. */

and this:

      u8      nonagle     : 4,/* Disable Nagle algorithm?             */

and looking into the kernel source code:

./fs/ocfs2/cluster/tcp.c:        ret = sock->ops->setsockopt(sock, SOL_TCP, TCP_NODELAY,

./net/rds/tcp.c:        sock->ops->setsockopt(sock, SOL_TCP, TCP_NODELAY, (char __user *)&val, the above are just two of the examples of how to set TCP_NODELAY in the kernel.

At the userspace level:

http://www.linuxquestions.org/questions/other-*nix-55/how-do-i-disable-the-n...

should be the correct way, except that the poster is attempting to do it in a kernel that does not support it.

As documented in http://linux.die.net/man/7/tcp, the supported kernel is 2.5.71 and onwards.

 Okay ... so, the situation I have at hand is that there is this app I cannot modify. I'd like to see its performance with nagle disabled. Is this possible?

as u can see from the above two kernel implementation - each is implementing the algorithm per-socket level, or connection-level, on top of TCP, so, yes, u want to do that is possible. but u will need to write a kernel module, that implement itself as a netfilter module, specifically just for the port that your application is using, and setup socket connection of TCP_NODELAY type. not sure if it will work?

While at it, is there a way to see the socket options of the socket descriptors in a running app? Regards,

Yes, i think so, just sniff out the network packet, and identify the correct bit as according to the header: include/linux/tcp.h: <snip> u16 advmss; /* Advertised MSS */ u8 frto_counter; /* Number of new acks after RTO */ u8 nonagle : 4,/* Disable Nagle algorithm? */ thin_lto : 1,/* Use linear timeouts for thin streams */ thin_dupack : 1,/* Fast retransmit on first dupack */ unused : 2; <snip>

Kashyap

-- Regards, Peter Teoh

frankly it is not wise to disable nagle, and the solution i proposed (using netfilter extension) is really too tedious - the added overhead may slow down the processing, even though disabling nagle aims to improve latencies. but modern network card is so fast, that the bottleneck is not at the network card side, but more likely in the network gateway, or router side - so u are definitely not going to see improvement in spite of all the troubles. -- Regards, Peter Teoh

On Tue, May 31, 2011 at 09:41, Anupam Kapoor <anupam.kapoor@gmail.com> wrote:

yup, that's what i also think. is it possible that nagle+delayed-ack is causing the perceived slow-down ?

could be, but have you tried to switch and use various congestion control ? check /proc/sys/net/ipv4/tcp_congestion_control and here are enabled ones from my distro stock kernel: $ grep -i cong /boot/config-`uname -r` CONFIG_TCP_CONG_ADVANCED=y CONFIG_TCP_CONG_BIC=y CONFIG_TCP_CONG_CUBIC=y CONFIG_TCP_CONG_WESTWOOD=m CONFIG_TCP_CONG_HTCP=m CONFIG_TCP_CONG_HSTCP=m CONFIG_TCP_CONG_HYBLA=m CONFIG_TCP_CONG_VEGAS=m CONFIG_TCP_CONG_SCALABLE=m CONFIG_TCP_CONG_LP=m CONFIG_TCP_CONG_VENO=m CONFIG_TCP_CONG_YEAH=m CONFIG_TCP_CONG_ILLINOIS=m CONFIG_DEFAULT_TCP_CONG="cubic" -- regards, Mulyadi Santosa Freelance Linux trainer and consultant blog: the-hydra.blogspot.com training: mulyaditraining.blogspot.com