[bind10-dev] Shared memory data source ideas

[JINMEI Tatuya / 神明達哉]

At Fri, 16 Mar 2012 09:33:30 +0100,
Shane Kerr <shane at isc.org> wrote:

> I tried to write up the ideas around a data source built around shared
> memory that we came up with last summer at the Amsterdam face to face
> meeting. I think I got the main points, but it has been quite a while
> and other people probably have different interpretations.
> 
> Anyway, they are here:
> 
> https://bind10.isc.org/wiki/SharedMemoryDataSource

We could also use an on-disk memory image (which would be mmaped).  It
will make the initial load time effectively zero.  I was not sure if
such an option is included in the above document.

In my understanding, updating an already established memory block is
generally difficult especially when it requires a lot of additions.
It was not clear to me from the document how we'd go about that.

Do we have one memory "block" for each zone or share a block for
multiple zones?

Not related to the main topic, but I don't think this is very accurate
in this context:

  One of the fundamental problems with BIND 9 performance has been
  locking data structures to allow concurrent access by multiple
  threads. This requirement for locking slows all lookups, and causes
  contention that limits the scalability of the program.

Authoritative-only BIND 9 named should be almost lock free if the zone
is "static" (not dynamically updated) and is as far as I know quite
scalable in that setup.  I know there is still a known upper limit of
number of effective cores, but it's less like due to locking issues.

---
JINMEI, Tatuya