Re: Cache Digest rejection %

On Tue, 20 Oct 1998, Niall Doherty wrote:

> The mgr:events shows that the peerDigestValidate starts fairly
> quickly after startup,

A peerDigestValidate is queued right after the first access to the
corresponding peer. The queuing delay depends on the number of peers that
are already queued for validation (so we do not fetch all the digests at the
same time)...

> but the rebuildDigest doesn't start for
> quite some time...

Store (local) digest is always built from scratch on startup after
store_rebuild completes. The in-memory copy is always identical to the
disk-resident copy except for a [short] rebuild period.

> am I correct in assuming the on-disk version is
> validated each time on startup to build the in-memory one ?

No. Old disk version is discarded when fresh digest is built. Digests are
always built in memory and then swapped out to disk for peers to fetch.

> So, it would mean that I can stop/start and be sure the same on-disk
> version is being used each time ?

I do not think so. You get a fresh digest after every startup. There is a
chance that a peer will re-use the old disk-resident copy _while_ the new one
is being built though. Not sure if it is a desired behavior. We could disable
disk-resident copy just before we rebuild the digest if that sounds better.

> With:
>
> refresh_pattern . 0 20% 2880
>
> Cache Digests histogram:
> Category Count %Total
> conf_max_age_stale 193470 81.91
> last_modified_factor_fresh 30702 13.00
>
> added: 32287 rejected: 203961 ( 86.33 %) del-ed: 0
>
> and with:
>
> refresh_pattern . 0 50% 43200
>
> conf_max_age_stale 4907 2.08
> last_modified_factor_fresh 169019 71.56
> last_modified_factor_stale 40442 17.12
>
> added: 170607 rejected: 65597 ( 27.77 %) del-ed: 0
>
> So - what's this telling me ? hm...

Hard to say. It all depends on what hit (and false hit) ratios you are
getting with each of these patterns. Play with them and see what works the
best for you... And yes, I know that try-and-see is not easy in this case.

Alex.
Received on Tue Oct 20 1998 - 13:36:13 MDT