memory partitioning

[dormando]

Peter Chiu wrote:
> My previous mail might be confusing.. pls let me clarify:
> 
>  
> 
> Option 1:
> 
> One instance of Memcached using 30GB of RAM running on each server. That 
> is, a total of 10 instances of Memcached
> 
>  
> 
> Option 2:
> 
> 15 instances of Memcached, each using 2GB RAM, running on each server. 
> That is, (a total of 15 instances on each server) x (10 servers) = 150 
> instances of Memcached
> 

The differences here are pretty obvious:

- Fewer large memcached's allow for higher efficiency of a get_multi.
- More instances mean you may logically partition them better. If you 
have datasets which vary widely in size it can be useful.

I'd default to fewer big instances and use threads, and set up anything 
else if you really know better.

-Dormando