[pox-dev] Very low performance in a database backed controller.

[Saul St. John]

On 05/13/2013 05:31 AM, Murphy McCauley wrote:
> On May 12, 2013, at 11:15 PM, Karthik Sharma wrote:
>
>> I have a function called act_like_switch which uses a databased 
>> backed architecture.If I come across a packet whose source address is 
>> not in the database I insert the <src_packet,port_no> into the 
>> database.Then I do a query of the destination address in the database 
>> and retrive the port_no from the database and then send out the packet.
>>
>> As you can see in the code below I measure the time taken for insert 
>> and query.I also record the counts as well.Apparently the average 
>> number of inserts is 30 per seconds and the average number of queries 
>> is 700 per seconds.This is very low for my application.Are there any 
>> ways to improve this numbers? Am I doing anything wrong?
>
> This doesn't seem totally outlandish to me.  That's like 33ms per 
> write... databases incur a penalty for their consistency properties 
> and so on.
>
> I don't know why you're doing this, so it's hard to evaluate your 
> design.  I'd suggest you ask yourself a bunch of questions... Do you 
> need to write this stuff to a database?  Do you need to do it this 
> often (could you store this in a Python data structure temporarily and 
> write it to the database every few seconds)?  Do you need to do it for 
> every exact-match flow (or could you do it for broader entries like, 
> for example, l2_pairs does)?
>
Also, instead of running a database lookup on each packet, could you 
push match rules into the switch? That way, most of your packets would 
avoid the overhead of being sent to the controller and waiting out the 
database query.

It's kinda the whole point of having flow tables, right?