The benchmarks now require "cargo bench --features=nightly". The
extra #[cfg(nightly)] switches in the code needed for it are a bit
annoying; I may move the benches to a separate directory to avoid this.
But for now, this works.
I should have submitted/pushed more incrementally but just played with it on
my computer as I was learning the language. The new Rust version more or less
matches the functionality of the current C++ version, although there are many
caveats listed below.
Upgrade notes: when moving from the C++ version, I recommend dropping and
recreating the "recording_cover" index in SQLite3 to pick up the addition of
the "video_sync_samples" column:
$ sudo systemctl stop moonfire-nvr
$ sudo -u moonfire-nvr sqlite3 /var/lib/moonfire-nvr/db/db
sqlite> drop index recording_cover;
sqlite3> create index ...rest of command as in schema.sql...;
sqlite3> ^D
Some known visible differences from the C++ version:
* .mp4 generation queries SQLite3 differently. Before it would just get all
video indexes in a single query. Now it leads with a query that should be
satisfiable by the covering index (assuming the index has been recreated as
noted above), then queries individual recording's indexes as needed to fill
a LRU cache. I believe this is roughly similar speed for the initial hit
(which generates the moov part of the file) and significantly faster when
seeking. I would have done it a while ago with the C++ version but didn't
want to track down a lru cache library. It was easier to find with Rust.
* On startup, the Rust version cleans up old reserved files. This is as in the
design; the C++ version was just missing this code.
* The .html recording list output is a little different. It's in ascending
order, with the most current segment shorten than an hour rather than the
oldest. This is less ergonomic, but it was easy. I could fix it or just wait
to obsolete it with some fancier JavaScript UI.
* commandline argument parsing and logging have changed formats due to
different underlying libraries.
* The JSON output isn't quite right (matching the spec / C++ implementation)
yet.
Additional caveats:
* I haven't done any proof-reading of prep.sh + install instructions.
* There's a lot of code quality work to do: adding (back) comments and test
coverage, developing a good Rust style.
* The ffmpeg foreign function interface is particularly sketchy. I'd
eventually like to switch to something based on autogenerated bindings.
I'd also like to use pure Rust code where practical, but once I do on-NVR
motion detection I'll need to existing C/C++ libraries for speed (H.264
decoding + OpenCL-based analysis).
I'm seeing what is possible performance-wise in the current C++ before
trying out Go and Rust implementations.
* use the google benchmark framework and some real data.
* use release builds - I hadn't done this in a while, and there were a
few compile errors that manifested only in release mode. Update the
readme to suggest using a release build.
* optimize the varint decoder and SampleIndexIterator to branch less.
* enable link-time optimization for release builds.
* add some support for feedback-directed optimization. Ideally "make"
would automatically produce the "generate" build outputs with a
different object/library/executable suffix, run the generate
benchmark, and then produce the "use" builds. This is not that fancy;
you have to run an arcane command:
alias cmake='cmake -DCMAKE_BUILD_TYPE=Release'
cmake -DPROFILE_GENERATE=true -DPROFILE_USE=false .. && \
make recording-bench && \
src/recording-bench && \
cmake -DPROFILE_GENERATE=false -DPROFILE_USE=true .. && \
make recording-bench && \
perf stat -e cycles,instructions,branches,branch-misses \
src/recording-bench --benchmark_repetitions=5
That said, the results are dramatic - at least 50% improvement. (The
results weren't stable before as small tweaks to the code caused a
huge shift in performance, presumably something something branch
alignment something something.)
* Changed README.md commensurately
* Add cameras.sql to .gitignore to not commit personal camera data
* Change CMakeLists.txt to explicitly refer to hand-built libevent dirs
There's a lot of work left to do on this:
* important latency optimization: the recording threads block
while fsync()ing sample files, which can take 250+ ms. This
should be moved to a separate thread to happen asynchronously.
* write cycle optimizations: several SQLite commits per camera per minute.
* test coverage: this drops testing of the file rotation, and
there are several error paths worth testing.
* ffmpeg oddities to investigate:
* the out-of-order first frame's pts
* measurable delay before returning packets
* it sometimes returns an initial packet it calls a "key" frame that actually
has an SEI recovery point NAL but not an IDR-coded slice NAL, even though
in the input these always seem to come together. This makes playback
starting from this recording not work at all on Chrome. The symptom is
that it loads a player-looking thing with the proper dimensions but
playback never actually starts.
I imagine these are all related but haven't taken the time to dig through
ffmpeg code and understand them. The right thing anyway may be to ditch
ffmpeg for RTSP streaming (perhaps in favor of the live555 library), as
it seems to have other omissions like making it hard/impossible to take
advantage of Sender Reports. In the meantime, I attempted to mitigate
problems by decreasing ffmpeg's probesize.
* handling overlapping recordings: right now if there's too much time drift or
a time jump, you can end up with recordings that the UI won't play without
manual database changes. It's not obvious what the right thing to do is.
* easy camera setup: currently you have to manually insert rows in the SQLite
database and restart.
but I think it's best to get something in to iterate from.
This deletes a lot of code, including:
* the ffmpeg video sink code (instead now using a bit of extra code in Stream
on top of the SampleFileWriter, SampleIndexEncoder, and MoonfireDatabase
code that's been around for a while)
* FileManager (in favor of new code using the database)
* the old UI
* RealFile and friends
* the dependency on protocol buffers, which was used for the config file
(though I'll likely have other reasons for using protocol buffers later)
* even some utilities like IsWord that were just for validating the config
This isn't as much of a speed-up as you might imagine; most of the large HTTP
content was mmap()ed files which are relatively efficient. The big improvement
here is that it's now possible to serve large files (4 GiB and up) on 32-bit
machines. This actually works: I was just able to browse a 25-hour, 37 GiB
.mp4 file on my Raspberry Pi 2 Model B. It takes about 400 ms to start serving
each request, which isn't exactly zippy but might be forgivable for such a
large file. I still intend for the common request from the web interface to be
for much smaller fragmented .mp4 files.
Speed could be improved later through caching. Right now my test code is
creating a fresh VirtualFile from a database query on each request, even
though it hasn't changed. The tricky part will be doing cache invalidation
cleanly if it does change---new recordings are added to the requested time
range, recordings are deleted, or existing recordings' timestamps are changed.
The downside to the approach here is that it requires libevent 2.1 for
evhttp_send_reply_chunk_with_cb. Unfortunately, Ubuntu 15.10 and Debian Jessie
still bundle libevent 2.0. There are a few possible improvements here:
1. fall back to assuming chunks are added immediately, so that people with
libevent 2.0 get the old bad behavior and people with libevent 2.1 get the
better behavior. This is kind of lame, though; it's easy to go through
the whole address space pretty fast, particularly when the browsers send
out requests so quickly so there may be some unintentional concurrency.
2. alter the FileSlice interface to return a pointer/destructor rather than
add something to the evbuffer. HttpServe would then add each chunk via
evbuffer_add_reference, and it'd supply a cleanupfn that (in addition to
calling the FileSlice-supplied destructor) notes that this chunk has been
fully sent. For all the currently-used FileSlices, this shouldn't be too
hard, and there are a few other reasons it might be beneficial:
* RealFileSlice could call madvise() to control the OS buffering
* RealFileSlice could track when file descriptors are open and thus
FileManager's unlink() calls don't actually free up space
* It feels dirty to expose libevent stuff through the otherwise-nice
FileSlice interface.
3. support building libevent 2.1 statically in-tree if the OS-supplied
libevent is unsuitable.
I'm tempted to go with #2, but probably not right now. More urgent to commit
support for writing the new format and the wrapper bits for viewing it.
Scott Lamb
Dolf Starreveld