[bind10-dev] Socket statistics (ticket2156)

[JINMEI Tatuya / 神明達哉]

At Fri, 31 Aug 2012 20:22:32 +0900 (JST),
fujiwara at jprs.co.jp wrote:

> 1. socket statistics items are hard to count.
[...]
> At first, I added counters where logging codes exist.
>     git b3d9dd1d5624efbd25f04b025ebc34e1bf533c26
> 
>     SOCKET_ASIODNS_FD_ADD_TCP,         ///< adding a new TCP server
[...]

> Now, I added some accept/open counters.
>     git 6c6ff24ac1eaf91473d79c765c5ccb05b899abdb
> 
>     SOCKET_ASIODNS_IPV4_TCP_ACCEPT,    ///< number of IPv4 TCP accept
[...] 
> Does this approach reasonable ?
> Or add some codes into asiolink and asiodns to count as listed ?

I thought I made my comments in the ticket, but anyway, these all
depend on what kind of statistics information related to
socket/network I/O we want to gather in the first place.

I don't know how useful BIND 9's socket related statistics information
is to operators.  If they find that particular set of counters useful,
it's generally better to provide the same information, and if it's
difficult due to API limitations we need to think about how to work
around that.  On the other hand, perhaps most developers just don't
care about these counters.  If that's the case it's probably just
better to skip implementing them in BIND 10, at least initially.
Or, people may rather want a different set of counters.  If so, we
should try to implement them.

For example, what about JPRS operators?  Do they think these
statistics counters of BIND 9 useful?  Or rather crucial?  Or not that
important?

> 2. where to hold counters ?
> 
> class isc::io_service::IOService seems to be a base class for socket
> related codes. May I add counters in it ?
> 
> Where do you think to add a socket statistics counter ?

I don't know yet.  I think it depends on specifically which counters
we want to support.

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