Steaming "180 stereo voices" sustained or not ?

[sbenno]

in this thread:
http://www.northernsounds.com/forum/...ad.php?t=34014

Rich Pell wrote:
600 stereo voices? I dont think so.. Your lucky to get 180 on a P4 2.6.

The problem is "streaming 180 stereo voices" is too vague.

Because it could be 180 voices active while the sampler is playing lots of short notes with sustain pedal down which would be released quickly and thus coming mostly from the sample heads in RAM, or you could have
180 voices streaming long (dozen of seconds) sustained notes.


Try the latter and tell me if you can achieve it without troubles with any sampler. (from a single disk, not a monster RAID array)

Looking at the polyphony counter and tell "it's streaming 180 stereo voices" is just too cheap.
It does not tell you the whole truth.

If you can play 180 stereo voices with all notes playing for dozen of seconds then its true "streaming 180 stereo voices".

The first is called "streaming 180 stereo voices, peak performance" and it is higher than the sustained performance because the sample heads (eg the first 0.5-2secs) are cached in RAM and do not need to streamed from disk thus lightening the strain put on the disk.

the latter "streaming 180 stereo voices, sustained"

Anyway I'd be curious about numbers and experiences in this regard from various people .


cheers,
Benno

[Rich Pell]

Well, i just finished testing a new piano in the Kontakt player , and though my test werent very un-scientific, about 160 stereo voices was max (for one the the NI techs as well ) this is considered pretty good performance, when using this way. Playing consecutive (about 1/2 sec. apart) 10 note block chords with sustain pedal down. Like a Listz cres.
I`m sure your right ,there are other ways of getting more voices to stream cleanly. I`m sure K2 can do even better..Best Rich

[kbaccki]

Playing consecutive (about 1/2 sec. apart) 10 note block chords with sustain pedal down. Like a Listz cres.
I`m sure your right ,there are other ways of getting more voices to stream cleanly. I`m sure K2 can do even better..Best Rich

The same 10 notes? If so, "OS cache" will affect performance considerably. You could either enable the "bypass OS cache" (which is a kontakt thing, not sure if it's available in the player), or do a more aggressive test. What I did was set up a MIDI file tp play distinct notes.

[Rich Pell]

The same 10 notes? If so, "OS cache" will affect performance considerably. You could either enable the "bypass OS cache" (which is a kontakt thing, not sure if it's available in the player), or do a more aggressive test. What I did was set up a MIDI file tp play distinct notes.

Yes, it is on the player as well and yes the same 10 notes, but i tried reandom chords as well. I havent testing with that on. Thanks for the info ..I`ll try.
What were the Specs and results of your test Kbaccki ? Best, Rich

[kbaccki]

Yes, it is on the player as well and yes the same 10 notes, but i tried reandom chords as well. I havent testing with that on. Thanks for the info ..I`ll try.
What were the Specs and results of your test Kbaccki ? Best, Rich

Kontakt 1.5.3 ran out of gas between 83 - 86 stereo voices. Best results were actually with "bypass OS cache" enabled -- initial streaming of samples would drop the top 6-7 notes, then subsequent runs would stream all notes smoothly when OS cache was disabled. I performed 56, 74, and 90 note key tests. My DFD settings that produced best results 72K pre-load, and 720K voice buffer. Interestingly, going below 650K or so on the voice buffer reduces polyphony significantly. voice buffer seems to be much more important than pre-load size. this was on a 1.8GHz PIV, 7200RPM 20GB maxtor, 1GB. Another interesting point: for the 90 note test I was streaming 3 instruments, but splitting those instruments over two physical drives (on seperate IDE controllers) did not produce higher than ~85 voices.

[sbenno]

hi guys,
thanks for the responses.
Rich, try to build a midi file that fires up all the notes in the chromatic scala (eg all 88 piano notes) in sequence (I'd say fire up 10 new notes each second) and then holds the notes for 20-30secs.
You will see you will get much lower performance due to the constant strain put on the disk.

kbaccki, did you use long sustained notes or shorter notes like Rich ?

cheers,
Benno

[Rich Pell]

Alright, I`ll try your test Benno. Actuallly, i found that bypassing the OS cache did nothing to improve the performance on my system.
However, I found that these DFD settings REALLY improved the overall performance for long pedal down sustained 10 note piano chords in rapid time.
This is in Kontakt player , Kompakt or Kontakt 1.53 . (Havent optimized DFD in K2 to really do a proper test)

Preload buf. 384 kb
Voice buff. .75 Mb
reserved voices 256
mem total 192 mb

Then something really important we found. Limit the voices on the front panel to 64 voices. I tried many diff. voice limits but 64 seems to work the best for most pianos...
With these setting i was able to stream up to 160-65 St. voice for long periods with the method described above.. I`m happy I finally found a setting that works well for my system Rich
P4 2.4 2GB ram
Seperate Seagate drives for Samples and OS on
Highpoint Rocket IDE controler card
HT and VM disabled

[kbaccki]

hi guys,
thanks for the responses.
Rich, try to build a midi file that fires up all the notes in the chromatic scala (eg all 88 piano notes) in sequence (I'd say fire up 10 new notes each second) and then holds the notes for 20-30secs.
You will see you will get much lower performance due to the constant strain put on the disk.

kbaccki, did you use long sustained notes or shorter notes like Rich ?

cheers,
Benno

Benno, your test is very similar to mine. I play a 56-note gliss over ~3 measures at 120BPM, so it works out to about 9-10 notes per second. What I did was create a 7-note gliss pf white notes only, then copy that over 8 octaves (this was done in SONAR). Then select all notes and stretch them out to about 15 measures, and add a sustain pedal on before the first note and then pedal off at the end. So it's playing sustained for ~30 seconds.

Why only white notes? For my test I used some variant of white grand (maybe one of the economy patches, which uses the same samples just less velocities, so it loads faster than the full blown patch). White grand only provides samples of the white keys -- so if you want to avoid double-counting playback of the same sample, like I did, I play back only white notes. When I went to the ~74 note test (or maybe it was 77 notes -- 56 + 21), I then added an economy patch of the scarbee rhodes playing around 3 octaves of white notes. Then with the ~90 note test I added 3 octaves of scarbee wurli. The glissandos for all instruments were offset slightly, but all notes were triggered within the first 3 measures (6 seconds), and sustained for 25-30 seconds.

Playing only white notes should avoid double hits on individual samples (for certain sample libs), and I'd recommend that since I don't know what sort of buffer optimizations you may inadvertently trigger by playing the same sample more than once. I suspect there is some bufer optimization. Having the 3 different sounds also allows me to distinguish artifacts between patches -- e.g. pan the rhodes to the left, piano center, wurli to the right, then any sample hickups/farts/pops/etc. are easier to hear.

One last thing: the note velocities were note randomized in any way (if there's even a way to do that), so in order to accurately test DFD settings when OS cache was enabled, I'd run the test once, restart kontakt, load a different large piano library (like white sister) in order to flush out OS cache, then run the test again -- a very time consuming process. Enabling "bypass OS cache" obviates the need to flush anything, but it seems that OS cache is still used during the sample pre-loading phase, even when "bypass OS cache" is enabled, though it is not used during actual streaming.

Last last thing: because I wanted to comparison test Kontakt with HALion 2 I played back the test as a MID file loaded into Win Media Player, then the MIDI output of that set to one of my Delta1010's, fed into the MIDI input of another Delta1010, then MIDI routed to Kontakt VST within the Console VST/DX host (http://www.console.jp/eng/). Ironically, I never even got around to comparing against HALion 2 since there's no comparative way to determine voice count in HAL, and I was happy with my Kontakt DFD tweaks.

- Keith

[Theodor]

600 voices with K2 is no lie ! I did a test as well today ( Had a really fast chromatic scale ascending and copied it around 10 times in a row on a piano patch )

I got 530 voices before K2 said : "there was a problem streaming the DFD voices" or something like that . Also ,that was the reason i could not take a snapshot ( because it focused on the Box )

The funny thing is that the error appears but the playback doesn't stop ! Now that's true power ! It just cuts a few notes . Kontakt rocks.

Amd @ 2.1 Ghz . 1.5 gb Ram . Seperate 120Gb 7000rpm . Sonar . When i did the test there was other stuff loaded in the memory , a bus send towards convolution etc. Msn open, the usual stuff

The question is, will Kontakt be able to play those voices when 20 different instruments are accessing the samples on a 7000 rpm drive in real time ? When you load a piano all the data is almost serial on the hardrive so it's not that hard accessing and streaming it.

But if there are too things on the ends of a big rope......Who knows

[Theodor]

I focused on the desktop instead and did the test again . Here is the snapshot ( not sure if it was the highest value )

IMAGE - K2 Poly maxed 78 Kb