Re: throughput and interrupts

From: Julian Elischer <julian_at_elischer.org>
Date: Wed, 16 Aug 2006 15:57:04 -0700
Peter Jeremy wrote:

>On Wed, 2006-Aug-16 09:59:22 +0700, Bachilo Dmitry wrote:
>  
>
>>Oh, it's natd. Now I see, but I just don't get it. I know that natd is not 
>>efficient but, as I've said, at home I have 9 or almost 10 MB/sec through the 
>>natd, while at this particular server I see only 3,7 MB maximum. I've tried 
>>now to turn all the natting off and tried to download a file and got like 9 
>>MB/sec, so it is natd who loads the system up.
>>    
>>
>
>natd runs in userland so every packet has to be pushed out to userland,
>processed and pushed back into the kernel.  The vast majority of the
>overhead is the userland/kernel transition so natd gives you a basically
>fixed pps rate.  Your throughput will vary depending on the packet size.
>  
>

in 6.1 there is an in kernel version of natd..

man ng_nat

>  
>
>>Someone advised me to use pf or ipnat, but I never did that before and heard 
>>that this nats have some limitations (like ipnat can't translate icmp packets 
>>or something).
>>    
>>
>
>Some time ago, I switched from natd to ipnat at work because the
>overhead was getting too much.  (I've also switched hardware so I
>can't give you direct performance comparisons).  I have found some
>problems with IPfilter in -stable when combining ipfilter/ipnat,
>stateful filtering and conditional NATing (ie a packet to B gets NAT'd
>to C only if it came from A).  (The same combination works in IPfilter
>3.x on Solaris.)  Normal filtering and NATing works OK.
>
>  
>
Received on Wed Aug 16 2006 - 20:57:06 UTC

This archive was generated by hypermail 2.4.0 : Wed May 19 2021 - 11:38:59 UTC