Performance of auth-only vs recursor.

[Kevin Darcy]

Sven Ingebrigt Ulland wrote:

>On Tue, May 09, 2006 at 04:35:53PM -0400, Kevin Darcy wrote:
>  
>
>>Sven Ingebrigt Ulland wrote:
>>    
>>
>>>On Tue, May 09, 2006 at 02:04:25PM +0200, Sven Ingebrigt Ulland wrote:
>>>
>>>      
>>>
>>>>Auth-only: 11,123 +/- 58 requests/second. (very small variation)
>>>>Recursor:   7,541 +/- 18 requests/second. (very small variation)
>>>>        
>>>>
>>>Um, that should be the other way around. The recursor outperforms the
>>>auth-only. Sorry about the mixup.
>>>
>>>      
>>>
>>Those numbers are so low, it's anyone's guess where the bottleneck is. 
>>58 qps works out to barely more than 5 million queries in the course of 
>>a day, and I have aging old Sparc Ultra 2s, almost a decade old, that 
>>can do way more than that. With your hardware specs, I think your 
>>numbers should be much higher.
>>    
>>
>
>Oh, those numbers are over 11 thousand queries per second for the
>recursor, and over 7 thousand qps for the auth-only. The numbers
>behind them are the standard deviations obtained from running ten
>consecutive tests, and show that there was very little variation in
>the tests.
>
>I'm suspecting some differences in code-paths between the two, that
>can be the cause for the differences in qps performance. It would be
>great if someone knew details on this.
>
Sorry, I misread it. The "+/-" showed up strangely in my font and threw 
me off.

The codepaths would indeed be different for fetching cached data versus 
authoritative data. But I wouldn't expect such performance differences, 
in the absence of other system/network factors. I agree someone who is 
more experienced in BIND 9 performance tuning should probably weigh in 
here. I've never dealt with sustained loads like that (we tend to buy 
lots of smaller servers and distribute/load-balance them, rather than 
concentrating a lot of load on a small number of larger servers).

- Kevin