Re: Networking: Lots of collisions?

[Fredrik Pettai <pettai%nordu.net@localhost>]

On Oct 8, 2010, at 11:12 PM, Fredrik Pettai wrote:
> Hi, 

Forgot to mention that its NetBSD/i386.

Looking at vmstat, I can see some things that stand out more than others:

# vmstat
 procs    memory      page                       disks   faults      cpu
 r b w    avm    fre  flt  re  pi   po   fr   sr f0 c0   in   sy  cs us sy id
 1 0 0 316088 158332  208   1   0    0   13   53  0  0  836 1763 1678 0  4 96

one of two processes are occasionally in the run queue and during that many 
page faults surface:

[...]
 0 0 0 316092 158336    0   0   0    0    0    0  0  0 1538 3006 3144 0  6 94
 2 0 0 316092 158328 1261   0   0    0    0    0  0  0 1576 3039 3032 8  8 84
 0 0 0 316092 158328    0   0   0    0    0    0  0  0 1527 2900 3120 0  3 97

> I just installed a netbsd-5-1-RC4 as a dns server, and I see a lot of 
> collisions:
> 
> pcn0 in       pcn0 out              total in      total out            
> packets  errs  packets  errs colls   packets  errs  packets  errs colls
> 39428897     0 22000706     0 5500180  39428897     0 22001100     0 5500180
>    3227     0     1892     0   474      3227     0     1892     0   474
>    3373     0     2060     0   514      3373     0     2060     0   514
>    3168     0     1926     0   482      3168     0     1926     0   482
> 
> Now, since it's running in VMware, one could guess that it's a underlying 
> problem (in VMware or maybe even in the physical infrastructure).
> But I also have virtualized Linux machines that are quite busy too, and they 
> don't show this kind of networking problem.
> (They run in the same VMware hardware)
> 
> Trying to do a tcpdump shows that the netbsd system doesn't handle that very 
> well either:
> 
> # tcpdump -i pcn0
> [...]
> ^C
> 5 packets captured
> 2585 packets received by filter
> 1726 packets dropped by kernel 
> 
> Doing it on the Linux machine works fine:
> 
> # tcpdump -i eth0
> [...]
> ^C
> 2844 packets captured
> 2845 packets received by filter
> 0 packets dropped by kernel
> 
> To that I might add that the servers doesn't have any typical CPU load etc.
> 
> # top -o cpu 
> load averages:  0.59,  0.65,  0.65;               up 0+12:32:18        
> 23:05:05
> 24 processes: 23 sleeping, 1 on CPU
> CPU states:  0.0% user,  0.0% nice,  2.0% system,  2.0% interrupt, 96.0% idle
> Memory: 306M Act, 2852K Inact, 6040K Wired, 7980K Exec, 117M File, 155M Free
> Swap: 256M Total, 256M Free
>  PID USERNAME PRI NICE   SIZE   RES STATE      TIME   WCPU    CPU COMMAND
> 3929 user    85    0    94M   91M netio     20:49  2.69%  2.69% [dns process]
> 
> Anybody else that has seen something similar? (in VMware?)
> Any hints on what to do to make the networking stack more optimized? It's 
> currently just the defaults.
> 
> /P