Subject: Re: IO Congestion Control
To: Steven M. Bellovin <>
From: Sumantra Kundu <>
List: tech-kern
Date: 09/11/2006 19:36:25
>> L_(i+1) = k*L_i + (1-k)*L

As Bill had remarked, it is difficult to understand how you could map
network congestion control algos in our case.  What is shown here is
EWMA technique which is not an algorithm but a mechanism to smooth out
transient spikes.

Congestion control comes in the picture only after we know that a
congestion has occur. In TCP this is identified  by lost packets. And
as we  noticed, IO congestion is affected by  three main factors: (i)
the ratio of dirty vs free pages, (ii) the effective service time of
the writes, and (iii) the nature of writes. Note: all these approaches
are generic and is not tied to any specific technology. What really
changes are the queuing and performance delay figures which are just
inputs to our framework.