Re: cat /proc/net/tcp takes 0.5 seconds on x86_64

[Eric Dumazet <dada1@cosmosbay.com>]

David Miller a écrit :
> From: Eric Dumazet <dada1@cosmosbay.com>
> Date: Thu, 28 Aug 2008 08:20:51 +0200
> 
>> But for route cache, it is probably doable since we added the
>> rt_genid thing in commit 29e75252da20f3ab9e132c68c9aed156b87beae6
>> ([IPV4] route cache: Introduce rt_genid for smooth cache
>> invalidation)
>>
>> If we add a hash table for each "struct net"
>> (net->ipv4.rt_hash_table), we then could do something sensible when
>> an admin writes to /proc/sys/net/ipv4/route/hash_size or at
>> rt_check_expire() time, if hash table is found to be full...
> 
> The synchronization and implementation is not a problem for
> the route cache, I implemented this eons ago.
> 
>> 3) In rt_check_expire(), adds some metrics to trigger an expand of the
>>   hash table in case we found too many entries in it.
> 
> This is the problem and why I didn't just commit the patch I had back
> then.
> 
> We could not define a reasonable way to trigger hash table growth.
> 
> GC attempts to keep a resident set of entries in the cache, and these
> heuristics are guided by the table size itself.  So if you grow the
> table too aggressively this never has a chance to work.

Maybe because of overcomplicated algos in net/ipv4/route.c, and
mixing "number of entries in cache", and "hash table size" things...

Fact is that nobody wants to have eight elements per hash bucket,
especially in case of DDOS.

> 
> You want to respond dynamically to traffic in a reasonable amount of
> time, but you don't want to get tricked by bursts of RCU effects.
> 

Right, but we also use process context processing instead of plain
timer soft irq things, so at least RCU effects should now be OK.

> We never came up with an algorithm that addresses all of these
> issues.

Could you give us the pointer to your previous work ?

Thank you