Re: [PATCH] net/ipv4/proc: Avoid usage for seq_printf() when reading /proc/net/snmp

[Paolo Abeni <pabeni@redhat.com>]

On 11/14/24 11:19, David Wang wrote:
> At 2024-11-14 17:30:34, "Paolo Abeni" <pabeni@redhat.com> wrote:
>> On 11/11/24 05:56, David Wang wrote:
>>> seq_printf() is costy, when reading /proc/net/snmp, profiling indicates
>>> seq_printf() takes more than 50% samples of snmp_seq_show():
>>> 	snmp_seq_show(97.751% 158722/162373)
>>> 	    snmp_seq_show_tcp_udp.isra.0(40.017% 63515/158722)
>>> 		seq_printf(83.451% 53004/63515)
>>> 		seq_write(1.170% 743/63515)
>>> 		_find_next_bit(0.727% 462/63515)
>>> 		...
>>> 	    seq_printf(24.762% 39303/158722)
>>> 	    snmp_seq_show_ipstats.isra.0(21.487% 34104/158722)
>>> 		seq_printf(85.788% 29257/34104)
>>> 		_find_next_bit(0.331% 113/34104)
>>> 		seq_write(0.235% 80/34104)
>>> 		...
>>> 	    icmpmsg_put(7.235% 11483/158722)
>>> 		seq_printf(41.714% 4790/11483)
>>> 		seq_write(2.630% 302/11483)
>>> 		...
>>> Time for a million rounds of stress reading /proc/net/snmp:
>>> 	real	0m24.323s
>>> 	user	0m0.293s
>>> 	sys	0m23.679s
>>> On average, reading /proc/net/snmp takes 0.023ms.
>>> With this patch, extra costs of seq_printf() is avoided, and a million
>>> rounds of reading /proc/net/snmp now takes only ~15.853s:
>>> 	real	0m16.386s
>>> 	user	0m0.280s
>>> 	sys	0m15.853s
>>> On average, one read takes 0.015ms, a ~40% improvement.
>>>
>>> Signed-off-by: David Wang <00107082@163.com>
>>
>> If the user space is really concerned with snmp access performances, I
>> think such information should be exposed via netlink.
>>
>> Still the goal of the optimization looks doubtful. The total number of
>> mibs domain is constant and limited (differently from the network
>> devices number that in specific setup can grow a lot). Stats polling
>> should be a low frequency operation. Why you need to optimize it?
> 
> Well, one thing I think worth mention, optimize /proc entries can help
> increase sample frequency, hence more accurate rate analysis,
>  for monitoring tools with a fixed/limited cpu quota.
> 
> And for /proc/net/*, the optimization would be amplified when considering network namespaces.

I guess scaling better with many namespace could be useful.

Please try to implement this interface over netlink.

Thanks,

Paolo