Commentary 10
"Queueing Model Based Network Server Performance Control"
By Lui Sha, Xue Liu, Ying Lu, and Tarek Abdelzaher

        In their paper, ^?ueueing Model Based Network Server Performance 
Control^? Sha, Liu, and Abdelzaher explain their method of combining 
elements from feedback control and queuing theory to produce a means to 
keep the timing performance of a network server as close as possible to 
the service level specification.  In network servers, controlling time 
performance has posed to be a difficult task.  The researchers claim 
that the key is to develop a model predictive control in which the 
queuing model^? online solution produces feed forward control, which, in 
turn, lengthens the feedback loop by keeping the system in a state of 
almost equilibrium.  The way their model works is that first, results 
from the queuing theory are used to compute the service rate needed to 
produce a desired average delay.  Secondly, server resources are 
allocated to achieve this rate, and lastly, a feedback control loop is 
used to compare/contrast the delay produced to the desired delay and to 
adjust the allocation of resources accordingly in order to get the 
produced delay more aligned to the desired delay.  After building their 
model with the use of an Apache server, Sha, Liu, and Abdelzaher 
demonstrate experimentally that their model produces results that are 
better aligned to QoS specifications.

        The greatest limitation of this paper was the fact that the model it 
describes is still in the prototype stage at the time of this writing. 
The authors note themselves that it is necessary for them to undergo 
more study on the subject matter.  Three years have passed since the 
time of this writing^?as any more work/study been done on this model? 
Another limitation of the paper is the fact that Sha, Liu, and 
Abdelzaher only compare their model with a system that implements solely 
a queuing model based feed forward control and with a system that 
implements solely a feedback control.  How does the model compare to 
other designs?  The fact that this paper does not answer this question 
leaves me a bit skeptical on the validity of its actual performance 
compared to other mechanisms.

        Although I have studied little in the way of queuing theory and 
feedback models, this combination of the two seems to be a clever 
solution to the problem of timing performance of network servers. 
Although Sha, Liu, and Abdelzaher prove what they can, I would need to 
see more proof of the success of such a model before I would actually 
put it to use.  Perhaps in the future we will see implementations of 
this model in the real world (I have yet to see any that I am aware of). 
  It would be interesting to track the progression of the development of 
such a model, and I am curious as to where this development stands today.