About performance of D-Bus

[Havoc Pennington]

Hi,

On Tue, Nov 4, 2008 at 6:09 AM, Jerome Philbert
<jerome.philbert at gmail.com> wrote:
> 100ms / 4.7ms = 21 messages only
> Moreover, this is the ideal case where the CPU has nothing else to do than
> sending messages ...
> In the reality, the CPU has lots'of things to do, so I could never send up
> to 21 messages.
>

On a CPU this slow, it isn't clear that libdbus is remotely usable;
you may need a dbus implementation (or other ipc system) that is
making different tradeoffs in the direction of efficiency. Or
worst-case, you may need to figure out how to avoid some of the ipc.

If you made a dbus library that:
* always blocked, thus eliminating the need to build DBusMessage
objects or maintain a message queue or deal with main loop
* was not thread-safe, thus eliminating locking
* was not robust against hostile peers, thus eliminating message validation
* did not ever do checks like dbus_return_if_fail thus eliminating more code
* was not safe against out-of-memory, thus eliminating even more code
* did not try to "multiplex" code modules with "path registration" or
"handler registration", thus killing more code and overhead

I think you could have a library that was dramatically smaller and
faster, fwiw. Basically by making it much less flexible and robust,
and just being careful about using it. I believe an implementation of
dbus could be pretty close to raw socket speed by making tradeoffs
like the above.

Current libdbus I'm sure can be made 30% faster or 40% faster or
something like that, given enough effort, but if you are only getting
21 messages now, then even a 100% speedup would only get you to 42
messages. If that still won't be enough, I would not dive into libdbus
optimization, I would figure out how to somehow use dbus less.

Again, we'd definitely love to see someone work on making current
libdbus faster.

Havoc