> > In my WCW99 presentation I showed data on cache sizes up to 8GB on
> > two-week NLANR request streams of 5-7 million requests. Now I'm
> > working with six 28-day NLANR request streams, each of which contains
> > between 10 and 35 million requests. I'm simulating caches up to
> > 32GB. My data show that LFU with an aging mechanism offers
> > substantially higher byte hit rates than LRU; for instance, on our
> > largest trace, an 8GB LRU cache gives you a 51% BHR whereas LFU gives
> > you 55%. I'll have a paper ready for submission in a few weeks.
...
Suggestions for increasing the realism of my work at reasonable cost
are very welcome.
I would gently suggest that an 8GB disk cache size is no longer very
interesting. However, I would be very curious to know if replacement
algorithms made a difference in managing main-memory caches. A couple
weeks ago I ran some LRU simulations of main-memory caches against the
Polygraph "uniform" model, and the hit rates are very low (e.g. 10%
byte hit rate for 60MB cache).
Cobalt has a few customers who have been generous enough to agree to
donate their log files for analysis. I have to anonymize these logs
but once that's taken care of I can release them.
gid
Received on Tue Jul 29 2003 - 13:15:58 MDT
This archive was generated by hypermail pre-2.1.9 : Tue Dec 09 2003 - 16:12:09 MST