Re: [lvs-users] questions about schedulers

To: " users mailing list." <lvs-users@xxxxxxxxxxxxxxxxxxxxxx>
Subject: Re: [lvs-users] questions about schedulers
Cc: Horms <horms@xxxxxxxxxxxx>
From: Joseph Mack NA3T <jmack@xxxxxxxx>
Date: Sun, 1 Feb 2009 11:45:33 -0800 (PST)
On Sun, 1 Feb 2009, Sebastian Vieira wrote:

> While the 'wlc' scheduler is perfect, i've found that once
> the traffic goes really high (tested with apache's 'ab') the director
> stops forwarding packets. This is around 4-5 thousand connections.


> My
> guess is that this is because of the IPVS table (size=4096) which
> doesn't get cleared because of persistence.

the "size" is the number of different buckets. The table 
size is only limited by the machine's RAM.

> So i've been looking around for another scheduler and one that seems
> good is 'dh'. I've done some tests with it and i get really push the
> site to its limits without packets being dropped or rejected.

wonder why dh has no problems

        Any ideas why wlc would have problems with large 
numbers of (persistent) connections while dh doesn't?

> One thing i know from the 'dh' scheduler is that the connection to the
> realserver is based on a hash made from the CIP.

that's how the sh scheduler works. The dh scheduler makes a 
hash of the URL (or the site name at least).

> Does this mean that
> in theory connections from different clients can all go to the same
> rs? If not, how are the client connections balanced?

for connections balanced on a hash of the CIP and you have 2 
realservers, then half the CIPs will go to one machine and 
half to the other.

> I've heard/read mixed opinions on persistence. My own opinion is that
> it works great, provided you trust all clients. Obviously this would
> only work in a LAN, not for an LVS put in front of webservers while
> being available to everybody (aka the internet). Am i wrong in this
> assumption?

well persistence is a dodgey kluge and lots of things can go 
wrong (which you can fix). I would suggest sh instead of 
persistence. I don't know if sh covers all cases, and people 
are happy to continue using persistence, despite its 

> In the past i've compiled the kernel to up the IPVS table size from
> standard 4096 to a multitude of that. Lately i've read in the HOWTO
> that this value should only be altered if i know more about IPVS than
> the developers. My ego isn't that big so i've left it alone ;)

good idea.

> Basicly my question would be better formed as:  if i deploy an LVS for
> a high traffic website that requires the client to be routed to the
> same realserver for as long as their session on the realserver lasts,
> should i still be using the 'wlc' scheduler in combination with
> persistence or is another scheduler better suited?

persistence is supposed to be independant of the scheduler. 
However I don't know why you're having problems with wlc.

Joseph Mack NA3T EME(B,D), FM05lw North Carolina
jmack (at) wm7d (dot) net - azimuthal equidistant map
generator at
Homepage It's GNU/Linux!

Please read the documentation before posting - it's available at: mailing list - lvs-users@xxxxxxxxxxxxxxxxxxxxxx
Send requests to lvs-users-request@xxxxxxxxxxxxxxxxxxxxxx
or go to

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