Performance and dropouts
platendoos adds heavy work to rekordbox (three neural networks, caching, stems) and takes work away (the quieter-in-the-background mods, faster analysis). This page is about keeping playback clean while it does.
The short version
- macOS: keep audiomem on. Before a set, close other apps if the log says
audiomem: N GB of swap in use. On a Mac that swaps heavily, set rekordbox's audio buffer size to 256 or 512 (Preferences > Audio). - Analyse before the set, not during: while a deck plays, a run takes one track at a time (
analyze_live=1) or waits (analyze_live=0). - Separate stems ahead for local tracks (the stems window): a track with stored stems needs no separation at a load. Analysed Spotify tracks (Spotify guide) skip the analysis at a load.
- External stems drive: an SSD in a USB case, not a cheap stick (Caches).
- Windows: see Windows notes for the GPU and
worker_priority.
What the mods do during a set
- Stems wait for the beat grid. While a track is in beat detection, stems hold back, so a deck's grid isn't late because of them (
betterstems_grid_first). - One analysis at a time while a deck plays (
analyze_live). - The models load one at a time at launch, about 6 s together in the background; before 0.3.0, loading them all at once could take up to a minute.
- Nothing is written to a slow cache drive while a deck plays.
betterstems_live_restcan leave the GPU free part of the time while a deck plays (off by default).
Clicks when the Mac is short of memory (macOS)
rekordbox's audio runs in blocks of 128 samples: 2.9 ms each (its audio settings; 256 and 512 are the other choices). A block that isn't ready in time is silence: a gap of 2.9 ms on every deck, more when the wait is longer. macOS reports each as an overload, and a recording of rekordbox's output showed one gap per overload.
Almost all of them have one cause: the audio thread waits for memory that macOS had put away. When other apps fill the memory, macOS compresses and swaps out what wasn't used for a few seconds. Measured on a 16 GB Mac with 6 to 8 GB in swap: 9 to 23 gaps a minute with one deck just playing, and more right after a deck load, which allocates and clears several hundred MB at once. What the audio thread was waiting for:
- rekordbox's FX history buffers (nine in ten): 214 buffers of 3 MB (654 MB) that rekordbox writes all the time, FX on or off, a new 16 KB page every tenth of a second per buffer.
- The playing deck's own audio: decoded into memory at the load, read minutes later.
- Rarely used parts of the program, now and then.
Plain rekordbox has the same gaps (with every mod off: 32 a minute and 10 per load under the test below), and none of the mods' own code on the audio thread can wait: no locks, no allocation, no logging.
audiomem (macOS) takes the cause away where it can: it pins the FX buffers that are being written, keeps the next 20 s of every loaded deck's audio in memory (4 s on a deck that isn't playing), and keeps rekordbox's code and data in memory. Measured with 3 GB of other memory in use on top:
| Gaps per minute, one deck playing | Gaps per deck load | Longest audio block | |
|---|---|---|---|
| Every mod off | 32 | 10 | 37 ms |
| All mods, audiomem off | 30 | 14 | 21 ms |
| All mods, audiomem on (two runs) | 6 and 0 | 1 and 0.1 | 9 and 4 ms |
What it can't cover: a jump further than the 20 s (a hot cue, needle search) and small rarely used objects. The log says when memory is short (audiomem: 6.5 GB of swap in use ...): close other apps before a set, and consider a buffer size of 256 or 512, since four in five of the waits are shorter than 5.8 ms.
The window standing still at a load is rekordbox clearing the new track's buffers on the main thread: about five buffers of 57 MB for a local file, built twice. With audiomem_lazy the buffers come from the system already empty and nothing is cleared up front: about 0.04 s instead of 0.12 s per load.
On Windows no dropout was found in any run, plain or with the mods (below).
The window standing still at a deck load
- macOS: with
audiomem_lazy, about 0.04 s instead of 0.12 s per load. - Windows: about 0.13 s, as in plain rekordbox, down from 0.35-0.4 s, thanks to
worker_priorityandquickdeck_merge(waveforms and grid in one redraw).
Less work while idle
framerate keeps the window at 30 fps unless a deck plays or you're using the computer; fontcache, screeninfo and midipoll (macOS) remove slow calls rekordbox makes all the time; spotifylog stops it rewriting its Spotify log several times a second.