LVS
lvs-users
Google
 
Web LinuxVirtualServer.org

Re: Re: tuning of LVS director for heavy traffic

To: khiz nms <khiznms@xxxxxxxxxxxxxx>
Subject: Re: Re: tuning of LVS director for heavy traffic
Cc: "'lvs-users@xxxxxxxxxxxxxxxxxxxxxx'" <lvs-users@xxxxxxxxxxxxxxxxxxxxxx>
From: Julian Anastasov <ja@xxxxxx>
Date: Tue, 29 Jan 2002 13:52:11 +0200 (EET)
        Hello,

On 29 Jan 2002, khiz  nms wrote:

> by default the route cache size max is 4096  on RH 6.2
> do the guys who run lvs director actually increase the size of the route 
> cache ..
> this is rh6.2 with kernel 2.2.19
>
> i got the foll post from a list which goes against change the value of the 
> max size
> "Please, try to leave gc_thresh and max_size at their default values,
> but to decrease gc_min_interval to 1 and to decrease gc_elasticity to
> 2,3 etc. It should help a bit. As extremal measure you may set
> gc_min_interval to 0 and/or to increase gc_thresh. It is better to leave
> max_size intact."

        Yes, I know about this posting from ANK. But if you really
talk with many destinations then try to increase the rows. It
speedups the lookup and prevents the CPU to be used for realoocations
of route cache entries - the hash function used is not perfect.
You can see the effect by using testlvs:

http://www.linuxvirtualserver.org/~julian/#testlvs

        Start it with the desired number of clients to see what
throughput you can expect (in one direction). The different
LVS versions (2.2/2.4) can show different results.

> Do we also need to play with
> /proc/sys/net/core/netdev_max_backlog

        Not sure for this. May be it depends on the used NICs.

> TIA
> Khiz

Regards

--
Julian Anastasov <ja@xxxxxx>



<Prev in Thread] Current Thread [Next in Thread>