On Sat, 6 Mar 1999, Bob Shustr wrote:
> We are seeing cache hit rates below 55% after 1 hour runs with --dhr 55p.
> During the test, polyclt --verb_lvl 2 listings suggest the hit rate is
> much lower than 55%. Isn't --dhr 55p maintaining a constant ratio?
Dhr 55% specifies constant _offered_ hit ratio. That is, the ratio achieved
on an ideal proxy that caches and keeps all cachable traffic. The offered hit
ratio is not reported on the console as it is constant. The offered hit ratio
does not depend on the proxy under test.
The measured hit ratio (reported on the console) may be lower if a proxy does
not cache some cachable objects (e.g., proxy is too busy) or if a proxy
purged some previously cached objects (e.g., to free disk space for new
traffic). The measured hit ratio depends, in part, on the proxy under test.
To debug what is actually going on and to double check that Polygraph is
not mistaken, use the --prn_false_misses option:
The Web page describes what the false miss is.
Finally, to see the total measured/offered HR statistics, us lr:
pail> ./lr -in /tmp/some.clt.log | tail | fgrep hits
3601.00 104463| p-meas:
hits: 51852 / 51968 misses: 42201 / 42085 errs: 0 left: 61 xacts: 94105
Measured HR = 51852 / (51852+42201) = 51852 / 94053 = .5513
Offered HR = 51968 / (51968+42085) = 51968 / 94053 = .5525
Sometimes, the situation is much worse. Here is a similar output from an
.... p-meas: hits: 25010 / 87386 misses: 133141 / 70765 ...
Measured HR = 25010 / (25010+133141) = 25010 / 158151 = .1581
Offered HR = 87386 / (87386+ 70765) = 87386 / 158151 = .5525
P.S. Constant offered hit ratio was designed for the bake-off to have similar
traffic regardless of the box under test (fairness issues). "Best effort
constant" _measured_ hit ratio mode is technically possible and makes sense
in some cases, but is not supported.
This archive was generated by hypermail 2b29 : Tue Jul 10 2001 - 12:00:05 MDT