[bind10-dev] more about cache (Re: cache effectiveness)

[JINMEI Tatuya / 神明達哉]

At Mon, 4 Mar 2013 08:50:38 +0100,
Michal 'vorner' Vaner <michal.vaner at nic.cz> wrote:

> > That might actually be an option if we can assume the "layer-1 cache"
> > covers the vast majority queries.  But it would depend on many things
> > whether this design is really effective, so I suspect we cannot be
> > sure at least without some actual experiments.
> 
> I guess it's interesting enough idea that we should do some kind of experiment
> to confirm or reject it. Do you have an idea how to make such experiment, other
> than the analysis of the data you did before?

To this end I guess we need something beyond analysis, such as
benchmark with prototype implementations.  For example, it should be
relatively easy to implement two versions of "L2 cache" on top of the
experimental "L1 cache" implementation of my own.  the L2 cache is
either in the same local process (like in the case of using threads)
or in another  process.  On cache miss or TTL expiration L1 asks L2,
and for experiments we assume L2 always knows the answer (so no other
external query will be needed).  We can then use queryperf like tool
changing the rate of L1 cache hit to see the difference of overall
performance.

---
JINMEI, Tatuya
Internet Systems Consortium, Inc.