Re: [Geopriv] review comments on draft-thomson-geopriv-held-beep-00.txt

On Mar 18, 2007, at 5:48 AM, Winterbottom, James wrote:

> Hi Andy one quick response now.
>
>
>>
>> 3) The justification for using BEEP is that multiple HTTP connections
>> are bad.  Why?  Many, many multi-tier servers work this way, with the
>> app server maintaining a pool of persistent TCP connections to its
>> back-end database.  I remain unconvinced that this BEEP binding is
>> anymore efficient for the purpose outlined than multiple HTTP
>> connections.
>>
>
> Our experience with using HTTP in the manner you describe in systems
> where requests can each have a different response time is that for a
> similar sized system using HTTP you get about have the peak rate of
> something BEEP-like. That is not to say that you cannot perhaps  
> tweak an
> HTTP system to improve performance, but I remain unconvinced that you
> can tweak to get a 100% increase in performance.
>
> The problem is of course dimensioning enough threads and ensuring that
> you don't get queuing of low-delay requests immediately after
> delay-tolerant requests thereby failing to meet QoS. You also need to
> avoid queuing multiple delay-tolerant requests to avoid starving.

If you are suggesting that TCP congestion control doesn't work with a  
high number of TCP connections, I'll let you argue that with the  
Transport folks... I'm not saying that is either wrong or right.

However, BEEP's channel multiplexing over a single TCP connection  
doesn't seem to solve the problem you are talking about.  It simply  
lifts the problem one layer up.  Perhaps your HTTP implementation  
needed some tweaking.

If you are saying that your implementation partitions the channels  
into sets of pools based on the characteristic of the query, that can  
also been done with HTTP/TCP pools as well.  And I'd note, the HELD/ 
BEEP draft does not mention that.

-andy

_______________________________________________
Geopriv mailing list
Geopriv@ietf.org
https://www1.ietf.org/mailman/listinfo/geopriv