[Tnfox-discussion] v0.80 of TnFOX released

[Niall D. <s_s...@ne...>]

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

Homepage: http://nedprod.com/TnFOX/
Docs: http://tnfox.sourceforge.net/TnFOX/html/

[Please note that FreeBSD and Win32 binaries will be uploaded 
tomorrow - sorry, I'm too tired to continue uploading tonight]

This is a big release for the third beta release of TnFOX. As would 
be expected in over three months of testing, a huge number of bugs 
have been fixed but furthermore features have been tweaked for easier 
usage, small new features added and some optimisation done on certain 
classes now that they are mature. The TnFOX homepage now contains a 
table showing the completedness of each of the TnFOX extension 
classes.

I'll only list new features here:
  * FOX v1.2.4 merged and FXString has been fully TnFOX-ised and now 
throws an exception if memory is not available for the string.

  * Lots of default constructors added to classes which didn't have 
them. Many other small API tweaks here and there.

  * FXIPCChannel is now mature having been very extensively tested 
and hand optimised:- New message routing facilities have been added 
and messages can now be tunnelled from one channel to another even 
ones unknown to the bridge; Messages now self-optimise endian 
conversion automatically; Extra robustness features are now present 
to allow FXIPCChannel to be attacked maliciously. Testing shows that 
FXIPCChannel is substantially faster than CORBA (OmniOrb).

  * All QDictBase subclasses can now have their table size be 
dynamically adjusted according to runtime factors by QDICTDYNRESIZE.

  * FXThreadPool now works, having been unintentionally completely 
broken in previous versions. It's also mature and hand optimised.

  * Generic::instantiateH has been reimplemented - it now causes less 
compiler slowdown and is much faster to compile. All metaprogramming 
tools can now take up to four variable parameters.

  * Build system has been much enhanced:- Whole Program Optimisation 
supported for compilers which have it (MSVC, ICC); Due to user 
request, build system can now always generate a static library; MSVC 
and ICC optimisation options have been hand tuned for performance.

  * Full support for GCC v3.5 upcoming symbol visibility extension 
added - now all private classes are marked as such, substantially 
improving generated code quality.

  * Linux kernel 2.6 support added. Kernel 2.6.6 is noticeably faster 
and less memory hungry than 2.4 which is good because 2.6.5 was slow 
and had broken shared memory too. All Linux precompiled binaries are 
now for RedHat Fedora Core 2.

  * Substantial improvements to the custom memory pool 
infrastructure, including a new port of ptmalloc2 from glibc v2.3.3. 
Threads can now set a "current heap" which prevents remote resource 
depletion attacks. This infrastructure is now mature and hand 
optimised.

  * Threading support is now mature and hand optimised. In 
particular, all operations involving a FXMutex are now as efficient 
as it's humanly possible to make it.

  * Lastly, unfortunately the python bindings library on POSIX still 
doesn't work fully - any exceptions thrown by the library (eg; when 
python raises an exception) lose their typeinfo should they exit that 
shared object and thus fail to match any catch handler. This may be a 
long-standing bug in GCC/GNU ld/Linux 
(http://sources.redhat.com/ml/libc-alpha/2002-05/msg00222.html) or 
indeed, it may not. Let's hope I can find a solution by next release -
 v0.80 is actually a whole seven days late just from investigating 
this one issue :(

A few more words on FXIPCChannel performance - on my machine (dual 
Athlon 1700 running Win2k), I can sustain the following figures for 
operations across two IPC channels across pipes with a tunnel on the 
remote end:

Ops per sec: 29770 asynchronous, 5306 synchronous.

Multiply the first by two and the second by four to roughly calculate 
ops/sec across a single connection. Pipes are known to be much slower 
on Win2k than Linux.

With sockets: 12981 asynchronous, 3018 synchronous. With sockets and 
128 bit AES encryption: 5633 asynchronous, 1469 synchronous.

All these figures can be improved upon again with Profile Guided 
Optimisation (ICC, the upcoming MSVC8) or even simply symbol ordering 
within the binary to improve cache locality.

Enjoy!

Cheers,
Niall





-----BEGIN PGP SIGNATURE-----
Version: idw's PGP-Frontend 4.9.6.1 / 9-2003 + PGP 8.0.2

iQA/AwUBQPNdyMEcvDLFGKbPEQKQZwCguQsryau3WYJMnkGSbzRO/Fu7Q28AoImF
cWHPqayORGcC0eIKitjWxess
=8rbt
-----END PGP SIGNATURE-----