The Planet Forums > RR loadbalancing is not even

mv_

Dec 22 2006, 12:29 PM

I know round robin load balancing does not provide any failover and even distribution of traffic, but I have strange situation - I have one website being served from 6 servers and one server gets 140mbps traffic in peak time while the other 5 servers get equal value around 84mbps. They are all on 1gbps ports, not shared by anything else.

I checked my dns entry several times and all is set exactly the same for all servers. These are all within one switch so from the dns server the 'distance' to other servers is the same.

Any ideas what I should check and achieve more even traffic distribution?

thank you

marian

Victor_n3v

Dec 28 2006, 05:33 PM

Hi mv_

Can you put what is domain name of the site so I can check everything from outside.

Good luck

Daroz

Dec 28 2006, 07:14 PM

Off the top of my head the biggest failure of RR load balancing is caching of the DNS entries.

Case in point, at one time (perhaps still) Comcast was accepting DNS responses with a very short (60s) TTL value and then caching them for 6+ hours. The net effect is that anyone using a Comcast DNS server would get the cached entry.

eth00

Dec 29 2006, 12:03 AM

That is why for most situations we do not suggest our clients use RR dns. There are many small issues that come up with it such as what you show. The caching of DNS can be a big deal, a few major ISPs caching will throw off the metrics.

As a semi-solution you may try to lower the TTL and if that does not help add a second A record for the server which is not getting as much traffic. Those combined may balance it out a bit.

mv_

Dec 30 2006, 01:36 AM

QUOTE (eth00)

That is why for most situations we do not suggest our clients use RR dns. There are many small issues that come up with it such as what you show. The caching of DNS can be a big deal, a few major ISPs caching will throw off the metrics.

As a semi-solution you may try to lower the TTL and if that does not help add a second A record for the server which is not getting as much traffic. Those combined may balance it out a bit.

I do this via A records, and my TTL is set to 240.

CODE

$ORIGIN .

$TTL 2400 ; 1 hour

domainname.com IN SOA ns1.mydns.com. ns.mydns.com. (

2006061007; serial

2400 ; refresh (1 hour)

600 ; retry (10 minutes)

86400 ; expire (1 day)

1200 ; minimum (1 hour)

)

$TTL 86400 ; 1 day

NS ns1.mydns.com.

NS ns2.mydns.com.

240 IN A 67.15.144.1

240 IN A 67.15.144.2

240 IN A 67.15.144.3

240 IN A 67.15.144.4

240 IN A 67.15.144.5

240 IN A 67.15.144.6

MX 10 mail.domainname.com.

$ORIGIN domainname.com.

ftp A 67.15.144.1

mail A 67.15.144.1

www 240 IN A 67.15.144.1

www 240 IN A 67.15.144.2

www 240 IN A 67.15.144.3

www 240 IN A 67.15.144.4

www 240 IN A 67.15.144.5

www 240 IN A 67.15.144.6

(those ips are fictional)

mv_

Dec 30 2006, 01:39 AM

QUOTE (eth00)

That is why for most situations we do not suggest our clients use RR dns. There are many small issues that come up with it such as what you show. The caching of DNS can be a big deal, a few major ISPs caching will throw off the metrics.

What would you recommend except hardware load balancer? I looked at solutions like lbnamed and I'm not skilled enough to test it on live busy website. Is there any solutions you imlemented that can handle lot of traffic? Some proxy server, or dedicated software load balancer? I really need it to redistribute traffic, no health checks are necessary.