What's wrong with the CSS ACA balancing method

[pcoughlin01]

I've read a number of posts that the ACA balancing method for the CSS is not recommended because you never really know what it's going to do. 

Is this because of it's high flows/second requirement to effectively balance equally, or is it just simply inaccurate? 

I've been using this method for many years without noticing huge differences in service hits.  Of course, it's possible that there's been a high number of flows/second to keep it accurate, although I don't see that in the flow stats. 

The primary reason I use this method is because it "appears" to take server response time into account.  All of the other balancing methods seem arbitrary regardless of the server health or performance.  The whole concept of load based upon the server performance just seems like the best bet. 

Wouldn't you want the method that has some sort of a measure of server perfromance even if it's not 100% accurate, or is ACA really that bad?

I'm curious to see who's using ACA out there.

[Daniel Arrondo Ostiz]

Hi,

There is nothing wrong with the ACA protocol itself, it's simply that it may be a bit unpredictable. If for example one of the servers is a bit slower than the rest, it will get less connections. Many people don't like this and prefer other simpler (but also effective) algorithms such as round-robin or leastconn

With that said, as any other protocol, it's supposed to give a good statistical distribution of the load if you look at a longer period. If it's working well for your setup, there is no reason to consider changing it.

Regards

Daniel

[pcoughlin01]

Yep, that's exactly what I'm seeing.  The longer it runs, the connections seem evenly distributed.  It can definitely be a pain when you're testing because you cannot determine what server will get the next connection, however in production use, we don't really care because the origin servers are equally configured.

I actually like the ability of the algorithm to not give new connections to a slower server.  All things being equal, with the origin servers configured exactly the same in terms of hardware, software, and application settings, then over time it ultimately evens out.

Do you see it in use a lot out there?

What about the ACE4710?  I've been using the predictor "response syn-to-close" to sort of mimic that same behavior whereby the load-balancer measures server response time.  How's that method?

thanks for your help.

Pat

[Daniel Arrondo Ostiz]

Hi Pat,

Most people in the field use either round-robin or leastconn, because both have very good load-balancing results while, at the same time, they are quite simple. However, there are no better or worse algorithms, you just need to choose the one you like most and the one that works best for your setup.

Daniel

[pcoughlin01]

Thanks for the info.

Pat