[vox-tech] Very slow TCP transfer over loopback

[Ken Herron]

--On Sunday, February 29, 2004 08:56:42 PM -0500 Mike Simons 
<msimons@moria.simons-clan.com> wrote:

> Hello All,
>
> Got a mystery here... transferring data over the loopback interface
> on a old RHAS 2.1 kernel (2.4.9-e.25enterprise) even a modern 2.6.1
> kernel, is _very_ slow.  When the SO_RCVBUF is set low.
>
>   As you can see from the tcpdump below after the session gets going,
> only one data packet is exchanged every .2 seconds, the sending side
> of the pipe seems to be waiting for something...
>
> - Anyone have a place to look in the RFCs or such that would explain
>   why it's waiting, for what?

Seems pretty quick to me:

Run while stracing the server:

> dd if=/dev/zero bs=1024  | time nc -q 0 localhost 2222
Command exited with non-zero status 1
0.07user 0.88system 0:38.67elapsed 2%CPU (0avgtext+0avgdata 0maxresident)k
0inputs+0outputs (144major+26minor)pagefaults 0swaps
363244+0 records in
363243+0 records out

Fixed the sample program to setsockopt() the accepted file and straced 
the server:

> dd if=/dev/zero bs=1024  | time nc -q 0 localhost 2222
Command exited with non-zero status 1
0.06user 0.79system 0:30.18elapsed 2%CPU (0avgtext+0avgdata 0maxresident)k
0inputs+0outputs (144major+26minor)pagefaults 0swaps
277564+0 records in
277563+0 records out

Run while tcpdumping the loopback interface:

> dd if=/dev/zero bs=1024  | time nc -q 0 localhost 2222
Command exited with non-zero status 1
0.17user 1.98system 0:21.79elapsed 9%CPU (0avgtext+0avgdata 0maxresident)k
0inputs+0outputs (145major+27minor)pagefaults 0swaps
818739+0 records in
818738+0 records out

Just let it run a while:

> dd if=/dev/zero bs=1024  | time nc -q 0 localhost 2222
Command exited with non-zero status 1
0.33user 4.90system 0:23.38elapsed 22%CPU (0avgtext+0avgdata 
0maxresident)k
0inputs+0outputs (144major+26minor)pagefaults 0swaps
2045878+0 records in
2045877+0 records out


-- 
Ken Herron