On Tue, 9 Apr 2002, Markus J. Kaiser wrote:
> i use polygraph against one squid server to extract the access.log
> file. the log file entries will be used against my caching
> algorithm for distr. proxy servers.
You can get even better quality logs if you use --dump hdrs option
with Polygraph client or server. Squid's access.log does not contain
enough information about message cachability and other properties that
may be important for your caching algorithm.
This way you can also eliminate the cache from the loop.
> in almost all polygraph tests i receive a declining hit rate (my algorithm
> is defined in such a way that it simulates the storage of 5000 objects,
> disregarding the object-size, on a hashing method as used in CARP).
>
> example :
> recurrence 0.6 & zipf(0.6) = graph declines for 1 mill. requests
> recurrence 0.5 & unif = graph declines for 700k requests
> recurrence 0.9 & zipf(0.9) = graph declines until 400k (HR=0.5) and jumps
> after 400k suddenly to HR 0.95.
> i received a similar declining graph with the polymix-1 file.
>
> it is not clear to me, why the HR is decreasing over time (i ran my simul.
> with different cache-sizes)
I think you answered your own question below.
> or why it suddenly increases in the last case
> (is there a "session" defined, when i dont define one? my test ran for 10
> hours)
Possibly because extremely high skew in URL access patterns can lead
to formation of temporary hot spots that increase hit ratio. You would
have to investigate your traces and, possibly, Polygraph source code
for the exact reason.
> when the reccurrence value is focusing on an over time increasing dataset
> how can the hit rate find a stable minimum (shouldnt it always decline?).
It cannot! That is why standard workloads such as PolyMix-4 limit the
working set size while allowing the working set to "slide" with time.
See, for example:
http://www.web-polygraph.org/mail-archive/users/200204/0001.html
BTW, PolyMix-1 is ancient and should not be used other than for
historical purposes. Working set size limits were introduced in
PolyMix-2. You should use PolyMix-4 as a reference point.
> is there a tool that allows me the extraction of the over-time hit rate (not
> the final value)?
Report Generator builds graphs using 5 second stat samples that
Polygraph logs (see --log option) and displays on the console (with
--verb_lvl 2 or higher). Report Generator uses lx and ltrace tools to
extract information from binary logs. You can write a custom tool that
uses console output or, better, ltrace output.
> the file with 1 mill. requests contains around 400k uniq calls ... what
> assumptions can i make in regard to the average number of stored objects
> (disregarding the object size).
Not sure what you mean here. Polygraph does not store any objects. You
can access Squid's storage statistics using its cache manager
interface.
Alex.
This archive was generated by hypermail 2b29 : Mon Feb 06 2006 - 12:00:22 MST