[Csci551-talk] mixed instructions from TAs

[S. Rubin]

Can the TAs synch up the responses, please, as we are getting some

mixed instructions.  For example, I posed a quesiton on Apr 20

re: testing of algorithms to be done with all clients starting at the same time.

Affan's responses (see below), implied that it was not a good idea.  I do 

not wish to re-run and re-write my tests and write-up due to the change in view

couple of days before the project is due, but I do not want to get penalized

either.  I ended up taking the mean of all clients' elapsed time and then

taking the mean of the means and the standard deviation at the end of each

run - more work, but a better estimate. Here is the trail of Q&A:

My question:  

2.  For phase 6, is it correct to assume that for
random algorithm all nodes start at the same time?
(Parts A and B state that about 2 other algorithms).

Affan's response # 1
> How do they say that? I have reread the specs but nothing seems to
imply that, and really the algorithm is localized, it shouldn't be doen
with assumtion of nodes starting at the same time.

My rebuttal:

Specs state the following for Part A:
 
First paragraph, last sentence:
"This algorithm should focus all peers on the same part of the file at
the same time if they start at the same time."
 
Then, the very next sentnce states:
"Question for README: how do you expect Algorithm S to perform
compared to Algorithm R assuming all nodes start at exactly the same time?
Why does it perform better or worse?"
 
and again for Part B, Question for README states the same thing.
Is it meant to be purely theoretical and not to be used for test cases?


Affan's response, which I took to mean "not a good idea":
Well you can use this condition to distill the behavior of the algorith,
Your program however should not *use* this as a neccecary condition. In
other words, simplify the infput to beter understand the algo, but donot
assume that we will (for checking) provide such simple input. 

Here is the trail from Xi:
It is specified that all nodes start at the same time. The purpose to 
test the system performance under heavy load.

For example if a new download tasks only starts when all previous tasks 
has completed the a peer-to-peer system would perform no better than a 
central server.

Narendra Acharya wrote:

> I am a bit confused about Xi’s answer. According to him we have to take 
> the time from start to the end of the completion of simulation. But this 
> works only if we consider that all the peers start at the same time. If 
> we consider different start time for different peers then we would not 
> be able to judge the performance of different algorithms. Therefore 
> isn’t it better if we consider the individual file download times for 
> each peer and then take required statistics? Because when evaluating the 
> algorithms we are not bothered about the time it take for a simulation 
> to complete, but rather are concerned about getting better file download 
> times.

-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://mailman.isi.edu/pipermail/csci551-talk/attachments/20050429/bda19b39/attachment.html