draining multiple sockets
Hi,
I am looking into the codes of linux kernel(ipv4/tcp.c,tcp_output.c /net/core/dev.c etc.) to find out how multiple sockets are drained on to the network. Here is the picture that I got so
far: If there are 'n' sockets with data to send over TCP, then they all supposedly take equal turns draining their data on to the xmit_queue.
1. However, this 'equal turn' (fair) philosophy is guided more in part due to the nature of the TCP behavior under competing flows, than anything else.
2. In practice though, the draining of sockets is like a FCFS scheduling discipline!!!
Which ever socket has data to send, will grab a piece of the TCP sk->sndbuf (16KB) and write it out. If the sndbuf is full, the process waits a random amount of time (between 2 and 21 jiffies) then retries again.
3. Once inside the sndbuf, the packet is passed on to the IP level and then to the device level (default device_xmit_queue length of 100 packets for ethernet). At every instant the routines constantly try to push as much data out on to the network (device_xmit_queue) as possible. right?
If I am wrong regarding any of the above 3 statements, **please** correct me.
socket1
||||| ---> TCP processing --> IP processing _
\ socket2 \ n/w xmit_queue |||| ---> TCP processing --> IP processing ------> |||||||---> . / . / . / socket n / ||||| ---> TCP processing --> IP processing -
- jaika
ps: Here is what I gleaned from the sources:
in tcp_ipv4.c: sk->sndbuff is initialised to sysctl_tcp_wmem[1]; and in net/ipv4/tcp.c: int sysctl_tcp_wmem[3] = {4*1024, 16*1024,
128*1024};
So I guess the sndbuff is initialised to 16K.
_________________________________________________________________
Get your FREE download of MSN Explorer at http://explorer.msn.com/intl.asp.
_______________________________________________
LARTC mailing list / [email protected]
http://mailman.ds9a.nl/mailman/listinfo/lartc HOWTO: http://lartc.org/
Quoted 6 lines · J K
2. In practice though, the draining of sockets is like a FCFS scheduling discipline!!!
Which ever socket has data to send, will grab a piece of the TCPsk->sndbuf(16KB) and write it out. If the sndbuf is full, the process waits a random amount of time (between 2 and 21 jiffies) then retries again.
16k=~10 packets, while there is a queue of 100 packets. So I think you will still achieve balancing.
The *real* experts on this reside on [email protected], you may want to ask there Alexey, Andi & DaveM are the real gurus.
Regards,
bert
Signature · 6 lines
http://www.PowerDNS.com Versatile DNS Software & Services http://www.tk the dot in .tk http://lartc.org Linux Advanced Routing & Traffic Control HOWTO _______________________________________________ LARTC mailing list / [email protected] http://mailman.ds9a.nl/mailman/listinfo/lartc HOWTO: http://lartc.org/