RE: content distribution in polymix-4 - some confusion

On Mon, 2005-09-05 at 14:30 +0100, Serge Ayoun wrote:

> Also since the average is measured at client side I do not see how byte
> hit ratio has any effect on the total avg size? Smaller object may be
> more popular, they still have an average which is fixed by the
> image/html/download/other object statistic.

Object content size distributions are indeed unaffected by object
popularity. Whether a given object (URL) is requested once or 100 times,
the size of that object would remain the same. In other words, all
responses to basic GET requests for a given URL should have the same
size. However, response sizes are measured for each response, not for
each unique URL. Thus, response sizes are affected by object popularity.

For example, requesting smaller objects more often will decrease mean
response size measured at the client because more responses will become
"smaller". To calculate true object/URL size distribution on the client
side, one would have to eliminate all repeated accesses to objects. This
can probably be done by enabling logging of URLs and response sizes.

Unfortunately, I cannot offer a formula to relate BHR discrimination
factor to mean object size skew. I am not even sure the effect of this
Polygraph algorithm is tractable. However, it should be possible to
build a graph that shows the dependency. 

Also, if everything is working correctly, then disabling BHR
discrimination should yield expected mean response sizes. This may be a
good starting point to understand what affect BHR has in your testing
environment.

HTH,

Alex.


> -----Original Message-----
> From: Alex Rousskov [mailto:[email protected]] 
> Sent: Sunday, September 04, 2005 7:02 PM
> To: Serge Ayoun
> Cc: users-mN9fGWdlm5pku/f+YMZH/di2O/[email protected]
> Subject: Re: content distribution in polymix-4 - some confusion
> 
> On Sun, 4 Sep 2005, Serge Ayoun wrote:
> 
> > It is written at
> >
> http://www.measurement-factory.com/results/public/cacheoff/N04/report.by
> > -alph.html#Sec:Polygraph that polymix-4 uses the following statistic:
> >
> >
> >
> > Images:            65%      4.5KB size
> > HTML:              15%      8.5KB
> > Download:           0.5%  300KB
> > Other:             19.5%   25KB
> >
> > Average of these is: A=10.575KB
> >
> > Also 85% are regular request and 15% are IMS. Among IMS, 66% requires
> > complete response.
> >
> > So calculating the final expected average response size, we should
> get:
> >
> > A*0.85 + 0*.15*.33+ A*.15*.66 = 10KB
> >
> > However in "Reply size" section, average size is said to be 6.8KB. How
> > is it possible?
> 
> Besides 304 responses to IMS requests (10%), there are several other 
> factors that decrease mean response size compared to the mean object 
> size. First of all, PolyMix-4 has a realistic byte hit ratio 
> simulation model (the PopModel::bhr_discrimination parameter), where 
> smaller objects are more popular than large ones. This leads to byte 
> hit ratio being smaller than object/document hit ratio, an important 
> characteristic of real traffic that earlier workloads were not able to 
> mimic.
> 
> There are also bodyless HEAD responses and aborted responses (0.2%).
> 
> HTH,
> 
> Alex.
>