|
>there was a thread here a while ago, where the design of load informed
>scheduling was discussed. An agent on the real-server monitors load
>(has modules for whatever you want to monitor) and sends info back to
>a demon on the director, which controls ipvs with ipvsadm.
Have you guys given any thought to the desirable frequency of such
load updates? This is esp. interesting in a high load situation...
Assume you have 50 real servers, and you are getting 5000 requests per
second.
Also, assume your real servers can handle max. 1000 HTTP requests per
second.
This means your load distribution can significantly change in as short of a
time
as 0.1 seconds. Even if we assume that my load scenario is exaggerated, it
is
possible to encounter peaks of this magnitude..
I would be interested in any comments
|
