Web Hosting Talk : Web Hosting Main Forums : Colocation and Data Centers : load balancing based on bandwidth throughput

[easygroove]

hello, are modern load balancers able to shift load between a cluster based on how busy the network card is?

To elaborate, say i ran a server farm which served mostly static content, i.e images, files etc etc, absolutely nothing dynamic that would max the CPU of the server, yet could reach the maximum capacity of the network card, would a load balancer be able to distribute the load between more than one server to keep the load equal?

Sorry if i missed the fundamental point of load balancing here but i always assumed that they shared load out based on CPU usage.

Thanks.

[astutiumRob]

Originally Posted by easygroove

Sorry if i missed the fundamental point of load balancing here but i always assumed that they shared load out based on CPU usage.

Depending on the type of LB you can balance
* equally across all available servers
* based on availability
* in a set preference
* based on cpu / disk / ram usage
etc

[Techee]

Remember that you can do simple load balancing between servers by using multiple IPs in your DNS records.

[easygroove]

Originally Posted by Techee

Remember that you can do simple load balancing between servers by using multiple IPs in your DNS records.

That's called a 'round robin' DNS am i correct?

Cheers!

[eth00]

Originally Posted by easygroove

That's called a 'round robin' DNS am i correct?

Cheers!

Correct, you also have very little control over it.

If you have a group of servers doing static content and all of them on the same size network card you could use just about anything for load balancing. There are plenty of hardware units out there that can do it or you can use one of the software options. Most will have different load balancing methods, for your case "least connections" would work fine. That would just send traffic to the server with the fewest connections and that should help keep the bandwidth about the same on each.

[iTom]

Originally Posted by easygroove

That's called a 'round robin' DNS am i correct?

Cheers!

Correct but it isn't great, you have no true control over the split of load.

Also if a server goes down there is no quick way to remove it from the cluster, due to DNS cache.

[lockbull]

CPU usage is rarely one of the metrics used for load balancing since it's not at the network level (you'd need to use an agent-based system on each server to do that). You'll find that most load balancers can distribute load based on bandwidth--you need to check the methods available for the device or software in question (these are generally called the distribution or load balancing methods). Aside from round robin, other typical methods are least connections, least response time (which would probably be the closest to something based on the servers' CPU usage without going to an agent-based system), least packets, etc.

[KarlZimmer]

You're probably just looking to do balancing based on the number of connections, which is a very common metric to balance by. The speed of individual connections and bandwidth utilization itself is not a very consistant metric from which to balance based on, especially if you look at it granular enough with all the microbursts that are happening. In theory, if you have enough connections and those are spread out equally, the bandwidth utilization will also be spread out relatively equally.

[swiftnoc]

Originally Posted by lockbull

CPU usage is rarely one of the metrics used for load balancing since it's not at the network level (you'd need to use an agent-based system on each server to do that).

Funny you mention this, its exactly what our load balancing system does. It installs an agent and takes into account whatever parameters are important for the client. (IE: connections, bandwidth, memory, I/O, CPU, type of service served from node, etc).
I believe that a good load balancer should take all this into account at the administrators discretion.

[funkywizard]

Originally Posted by iTom

Also if a server goes down there is no quick way to remove it from the cluster, due to DNS cache.

True, but in my experience, web browsers are smart enough to connect to one of the other ip addresses if it notices that the one it wants to connect to is down, with round robin dns. I certainly don't rely on this feature, since it's totally up to the browser to do this, but in my experience, if one server goes down, the other seems to pick up almost all the load from the downed server, rather immediately.

[plumsauce]

Originally Posted by funkywizard

True, but in my experience, web browsers are smart enough to connect to one of the other ip addresses if it notices that the one it wants to connect to is down, with round robin dns. I certainly don't rely on this feature, since it's totally up to the browser to do this, but in my experience, if one server goes down, the other seems to pick up almost all the load from the downed server, rather immediately.

Well, the disadvantage is that the browser still has to wait for a timeout before trying again. The site owner probably wants to avoid this.

[funkywizard]

Originally Posted by plumsauce

Well, the disadvantage is that the browser still has to wait for a timeout before trying again. The site owner probably wants to avoid this.

Definitely. I'm just saying that RRDNS is a bit more robust than I had assumed when I first started using it.