flss - Fast lean Streaming Server

Implementation status

The following table summarizes the status of the various project components

Item	Implementation Status	Last change	Comment
RTSP	in progress	2003-02-06	implementation is interoperable with various players (see below), but no rigorous standards compliance testing yet; only supported commands at this point are DESCRIBE, SETUP, PLAY, PAUSE and TEARDOWN
SDP	complete	2003-02-06	SDP files are generated on the fly for requested media; the format is understood by all tested players and fully conforms to RFC2327; a more abstract representation of media descriptions inside the server is required to improve on the current state
RTP	mostly complete	2003-02-06	unicast only
RTCP	in progress	2003-03-14	sender reports and sdes packets are periodically generated; reports from receivers are collected, but currently ignored
MPEG-1 Systems	mostly complete	2003-02-06	MPEG-1 elementary streams multiplexed in MPEG-1 system streams can be streamed natively; synchronization between video and audio tracks is not completely perfect when not starting playback from the beginning of the clip (no skew, but a small offset); total playtime calculation is not very accurate
MPEG-1 Video	mostly complete	2003-02-06	MPEG-1 video bitstreams can be streamed natively; random seeking into the stream is not very accurate, but essentially works as expected
MPEG-1 Audio	mostly complete	2003-02-06	MPEG-1 layer I, II and III files can be streamed natively; server currently aborts when encountering mp3 tags, they should be ignored; all audio streams are silently assumed CBR; while the server can send VBR streams without any problem, seeking and total playtime calculation are impossible in this case
MPEG/RTP encapsulation	complete	2003-02-06	fully compliant to RFC2250 and tested with various players; not all encapsulation options of RFC2250 are used yet
Run-time configuration	finished	2003-05-02	Configuration file has a C-like syntax, using bison for the parser; still need to expose more configuration variables through the config file
IO subsystem	in progress	2003-02-27	disk IO subsystem is in working state; disk bandwidth is kept track of, but no reservation is performed at the moment; similiar for cache memory
logging	in progress	2003-02-27	logger class definitions and implementation for stderr logging; logging functionality not used very consequently throughout the server
autoconf	complete	2003-05-02	finally found time learning autoconf; introduced into the project, and rewritten parts of makefile and config files accordingly; successfully integrated FreeBSD and Mac OS X ports
"code-lifting" (code beautification)	in progress	2003-02-27	canonify exception handling and throwing; add copyright headers to all files; clean up build system

Compatibility testing matrix

The following table lists the players which the server is currently being tested against; if you know of another player that might work, just mail me and I will include it in future test. Not all problems listed in the table below are necessarily due to bugs in the server.

Item	QuickTime Player	RealPlayer	mplayer+live.com
RTSP	ok	ok	ok
RTP	ok	ok	ok
MPEG-1 audio	ok (QT accepts only layer I, II)	ok	ok
MPEG-1 video	ok	player crashes	ok
MPEG-1 audio and video	ok; some curious problem with RTP-Info header field	player crashes	player has serious sync problems

Changelog

2004-08-13 Almost a year since I last touched the code... performance and portability is impressive, the server core is simply stable; I think I will just announce it in its current shape as it basically does what I have wanted it to do for the first release. Too lazy to make individual changelogs, so just a quick overview:
- Heavy performance instrumentation and tuning on all supported platforms; performance measurements (see performance evaluation).
- passes valgrind, electric fence
2003-08-12 Whoaa -- These pages are *way* out of date; lots of fixes and cleanups, and now seriously busy with performance-tuning:
- event dispatcher can now optionally use epoll on Linux, which is a significant performance improvement over poll-based dispatching
- fast atomic refcounters on ppc (formerly only available on i386)
- no memory allocations in fast path anymore
- can probably handle around 320 clients, MPEG-1 at 1.5MBit on a dual P3-933MHz
Discovered that the string implementation in STL shipped with gcc-2.95.3 is not thread-safe -- this means you have to use either gcc>=3.0 or stlport.
2003-05-05 finally a release!
2003-05-02 long time no update to this page, and lots of other things to do besides the server... server has been successfully ported to FreeBSD and Mac OS X, a lot of internal cleanup etc. No new features, and definitely moving towards a release
2003-04-09 almost compiles on OS X, need to implement getopt replacement
2003-04-03 autoconf, finally!
2003-03-31 gcc-3.x cleanups, fixed a severe bug in the block caching layer (I wonder how it was working at all before?!?)
2003-03-14 lots of code cleanup; now got a working config file parser, the server initialization phase is getting into shape
2003-03-13 spent half of the day chasing and fixing a subtle wakeup race in the core event dispatching mechanism; beginning to become convinced of the usefulness of formal program proves
2003-03-12 not very productive today; started sketching out RTCP finally; ran the server with efence for the first time (it survived! but have to set EF_ALLOW_MALLOC_0 because some exceptions thrown are "empty")
2003-02-27 finally managed to put up this webpage summarizing the present state

TODO list

convert all remaining functions from returing error codes to throwing exceptions (where appropriate)
instrument the program with valgrind and/or electric fence
general code cleanup, stylistic changes (some class and function names are not to my liking)
portability: investigate why mac os x is behaving strangely
"true" async io, start with posix aio (last time I tried something with librt on Linux two years ago, the threading lib was too buggy, hope this has been fixed in the meantime...)
increase logging verbosity -- turns out that removing the debugging cruft was not a good idea to do before porting to other platforms...
unit tests -- write unit test for the individual modules; I *wanted* to do this from the beginning, now I pay for laziness..