load balancing algorithms

[Maxim Dounin]

Hello!

On Tue, Sep 26, 2017 at 09:43:37AM -0400, Adam Schwartz wrote:

> > On the other hand, there is API in nginx which allows to implement 
> > any load balancing algorithm needed.
> 
> Cool! I was looking for something like that.
> 
> > As for the algorithms you've mentioned, "power of two choices" seems 
> > to be better than random, though it does not look like it is 
> > beneficial even compared to round-robin.
> 
> Are there specific reasons or example you have that show round-robin is sufficient?

As long as we are assuming single load balancer and identical 
requests, round-robin essentially means that we place equal load 
on all balanced servers, and this is better than what "power of 
two choices" can provide.  Certainly things will be different if 
requests are not equal, though this is what least_conn is expected 
to address (and again, it does so better than just testing two 
choices).

As far as I understand, "Power of two choices" might be beneficial 
if you need to optimize time spent on comparing the load on 
different balanced servers, for example, when balancing very large 
number of servers.  Or when trying to do distributed balancing 
with multiple load balancers, and query loads directly from 
balanced servers - but this is not something nginx currently tries 
to address.

-- 
Maxim Dounin
http://nginx.org/