Insanely great performance on ARM-64bit (Pianoteq, Odroid N2+, Linux) (Page 1 of 3)

A COMPARISON:

This is Pianoteq at maximum settings, run without being assigned to any cores. MIDI file is un-playable. Performance Index drops to 20 and there are red lines galore. *notice the audio load graph, the blue Performance Index line is a third of the way down from the top, orange audio load is very high and is consistently above blue Performance Index, Pianoteq reported CPU usage is a very high 76%.

Pianoteq with same maximum settings, but now, dedicated to 4 cores. The MIDI file plays smoothly, Performance Index is stable at ~34 all the way through the song (no wild fluctuations).There were only  2-3 red red lines (not shown). *notice the differences in the audio load graph, blue Performance Index line is pegged to top, orange audio load is down and consistently below the blue Performance Index. Pianoteq reported CPU usage also down to 37%.

Cool, thank you.

Here's the simple test:
- set everything in Pianoteq to it's highest setting
- without assigning any cores, play a MIDI file and take note of the performance index + graph
- then, re-start Pianoteq with assigned cores and play the MIDI file see how the performance index + graph are affected.

Performance is the same, I can't hear any quality differences between the onboard N2+ DAC and my external DAC. They both sound great with Pianoteq. The main difference: onboard DAC uses 3.5mm stereo jack, my external DAC has balanced XLR outs.

Specs:
The N2+ DAC/sound chip runs at 32-bit up to 192kHz.
My external DAC runs at 24-bit 96kHz

In Pianoteq:
The N2+ DAC/sound chip shows a bit depth of 32-bits up to 192000 Hz sample rate.
My external DAC runs at 24-bits up to 96000 Hz.

available options 44.1kHz and above:
44100 Hz (both)
48000 Hz (both)
88200 Hz (both)
96000 Hz (both)
176400 Hz (onboard DAC only)
192000 Hz (onboard DAC only)

For reference my external DAC is the Mackie MDB-USB DI box (has 2 XLR outs on the other side).

Is this what you're looking for?

USING ONBOARD N2+ DAC

USING EXTERNAL DAC (Mackie MDB-USB DI)

Comparison of CPU Loads with and without assigning Pianoteq to specific cores

WITHOUT Core Assignment, the default CPU load-balancing algorithm tends to:
1. Favor the first core, often overloading it while under utilizing the other cores.
2. It does eventually round-robin and distribute the work load across the cores, but it's too slow for real-time music. Many times, by the time it realizes there's been a stressful "musical" event, it's too late to do anything about it.
3. To add insult to injury, it's favoring the first 2x low-performance cores, and under-utilizing the 4x high-performance cores.

Notice:
- Pianoteq reads CPU load around 60%
- Performance Index dropped to 20
- There's a couple of CPU overloads/red x-runs and orange audio load consistently above blue Performance Index.

This is what it typically looks like when Pianoteq is assigned to specific cores (CPUs 3, 4, 5, 6).
- There tends to be a more even distribution across high-performance CPUs (3, 4, 5, 6)
- The 2x low-performance cores (1,2) aren't so overloaded and are free to attend to the needs of the OS and other programs.
- Pianoteq reads the CPU load as a more normal 302%
- Performance Index is pegged at 34 (no wild fluctuations), no x-runs/red cpu overloads, orange audio load is consistently below blue Performance Index.

Please keep in mind, these are snapshots of a live dynamic system, but I think these do show 'on average' CPU loads for each situation.

That’s a good point about On Demand vs. Performance CPU settings. It does look like the system is running ‘On Demand’ when no cores are assigned to Pianoteq. But when the cores are assigned to Pianoteq, those 4 cores act more like they are in Performance’ mode.

I saw that you have chosen NY Steinway D Player clean and
NY Steinway Model D in your tests. With two virtual mics these are not the hardest challenge.
In the Pi400-thread me and navindra tested with NY Steinway D Classical Recording (4 virtual mics and thick ambience), unmodified preset and Pianoteq's Blues Demo.

With Player Clean (following snapshot) and also Model D the Pi400 is not bad too:

At the same time my Pi400 drives a vncserver and sends the Pianoteq GUI data per WLAN to the client. The interface in the snapshot had been the 48kHz-USB-DAC in a Korg B2,
1000 Hz kernel,
PiOS 64-bit beta,
governor performance,
PTQ Standard v7.2.0.

I agree Armbian had very good performance with Pianoteq, it's a very nice light-weight linux distribution. I really like their kernel build for customizing. So far though all the linux distributions I've tried have run Pianoteq very well.

I downloaded several linux versions specific to the Odroid N2+ / ARMv8 architecture and bought extra microSD cards so I could test them out (honestly, so I could geek out, but also to test them with Pianoteq). Right now, I'm running almost all of them off individual microSD cards. When I finally settle on one, I'll flash that one onto the slightly faster eMMC card.

1. Ubuntu MATE
2. Ubuntu Minimal
2. Armbian (Desktop/FOCAL)
3. Manjaro ARM
4. Arch Linux ARM (A.L.A.R.M!)

Ubuntu MATE & Ubuntu Minimal
Since they are "officially supported" by Hard Kernel, it was easy to get the N2+ hardware running on both of them. Pianoteq installed and ran immediately on both and they installed with a low-latency PREEMPTive kernel. My linux skills are rusty, so I'm mainly using the Ubuntu MATE with the GUI to get up to speed. My favorite so far is Ubuntu Minimal; command-line only, but I have Pianoteq running nicely with the i3 Window Manager, nice clean setup. Responsiveness of Ubuntu Minimal is better, but not sure if it makes a significant difference for Pianoteq. After a bit of tussling and getting up to speed, I like both just fine. Hardware requirements for Ubuntu MATE 2GB RAM / 8GB of disk space. Ubuntu Minimal 1GB RAM / 4GB disk space. Ubuntu MATE will run more comfortably on the 4GB model of the N2+, Ubuntu Minimal is a good fit for the 2GB model.

Armbian
Very responsive and light distribution. I installed the Desktop/Focal version. Default is a PREEMPTive kernel. Pianoteq installed and ran immediately. Main issue - couldn't the get onboard sound / 3.5mm jack to work. Otherwise, Pianoteq played fine through HDMI and external sound card.

My understanding is that Armbian is built off the Hard Kernel Ubuntu distributions with different options turned on/off, but I haven't dug too deep yet. AND - I really like their helpful Youtube videos and kernel builder platform.

Manjaro ARM
Very slick with good support. It's a well-packaged, light-weight, responsive Arch based distribution. Pianoteq installed and ran immediately, but once again - no sound thru 3.5mm jack. Pianoteq worked fine through HDMI and my external audio interface. Unfortunately, default kernel is a NON-PREEMPTive kernel - so no low-latency out of the box. One of Manjaro's main selling points is the ability to select and switch to low-latency and real-time kernels, but this feature wasn't working yet on Manjaro ARM. I will re-visit when this is working properly. One thing I liked in Manjaro ARM, it offers a command-line installer called Architect, let's you install exactly what you want, no more, no less.

Arch Linux ARM (A.L.A.R.M.)
I don't have the technical chops to install this yet (mostly it's getting the boot card formatted properly). Manjaro bills itself as the user-friendly version of Arch, while Arch's philosophy is DIY - do/only install what you want. At any rate, there is an Odroid N2+ version of Arch.

What improvements do you get with the 1000 Hz kernel? Is that specifically for MIDI hardware or does it also affect MIDI file playback?

That's cool. Setting the system timer to 1000Hz is on my to-do list, but first, I have to figure out how to re-compile the Ubuntu kernel!

Note: I did find out that the default setting is 250Hz for Ubuntu-MATE on the Odroid N2+.

On advice from the Linux Musician forums I downloaded/ran the Perl scripts which check for real-time music optimizations.

Scripts can be downloaded here:
https://github.com/raboof/realtimeconfigquickscan

This runs the scripts in the Terminal

Results for my system.
I've already made many of the recommended adjustments so results are mostly "good". When a "not good" result is given, there is usually some advice how to fix it, for example the Kernel with Real-Time Preemption check below.

NOTE: by default, all the CPU Governors are installed as "performance" on Ubuntu/Odroid N2+, so that's nice.

This may not be needed in the case of the N2+, by dedicating Pianoteq to the 4 high-performance cores, it leaves the 2 low-performing cores for the OS. I tried assigning Pianoteq to 5 and 6 cores, but it seems there is something optimal about using only 4, which works out for the CPU time needed by the OS.

Yes, because this is the function:
T = 1/250Hz = 4ms (=4000.000µs system timer)

Minimal distros (without desktop gui) can be used, when the Pianoteq-GUI is started directly into a vncserver for example --> https://forum.modartt.com/viewtopic.php...11#p971611

MUSICAL STRESS TEST - #2

MIDI File: Étude in A Minor , Op. 10, No. 2, “Chromatic” by Frédéric François Chopin. Performed by Sijing Ye.
https://www.yamahaden.com/midi-files/ca...1667694550
Preset: 'NY Steinway D Classical Recording'
Settings: Maximum (first 5 seconds of video show the settings)
Results: Clean playback. Ignore red line at the beginning, which happens sometimes when I switch windows.

https://www.youtube.com/watch?v=WOu1SUL9ld0

* Set Youtube playback resolution to 720p or higher to see the graph more clearly. I used my iPhone to record the video and it's awful at low resolutions. Audio is from the 3.5mm jack, routed through an audio interface (SSL2) to iPhone.

MUSICAL STRESS TEST #4

MIDI File:
Étude in G-Sharp Minor , Op. 25, No. 6, “Thirds” by Frédéric François Chopin. Performed by Jean-Selim Abdelmoula
https://www.yamahaden.com/midi-files/ca..._490225942

Preset: 'NY Steinway D Classical Recording'
Settings: Maximum (first 5 seconds of video show the settings)
Test Results:
00:45 - minor event (single red line)
00:54 - minor event (single red line)
00:58 - minor event (single red line)
01:08 - minor event (multiple red lines)
01:22 - minor event (single red line)

https://www.youtube.com/watch?v=xhDEUw4u_SY* Set Youtube playback resolution to 720p or higher to see the graph more clearly. I used my iPhone to record the video and it's awful at low resolutions. Audio is from the 3.5mm jack, routed through an audio interface (SSL2) to iPhone.

MUSICAL STRESS TEST #6

MIDI File:
Étude in C Major , Op. 10, No. 1 by Frédéric François Chopin. Performed by Maria Verbaite
https://www.yamahaden.com/midi-files/ca..._863189919

Preset: 'NY Steinway D Classical Recording'
Settings: Maximum (first 5 seconds of video show the settings)
Test Results: 
00:07 - minor event (1 red line)
00:43 - minor event (2 red lines)
00:54 - minor event (4 red lines)
01:15 - minor event (1 red line)

https://www.youtube.com/watch?v=eHpvlri5jC0* Set Youtube playback resolution to 720p or higher to see the graph more clearly. I used my iPhone to record the video and it's awful at low resolutions. Audio is from the 3.5mm jack, routed through an audio interface (SSL2) to iPhone.

The 7 Musical Stress Tests I posted above are the easier/short ones. I tested another set of 7 more challenging pieces that were much longer (for example The Hammerklavier, Gaspard de la Nuit, and 3 Movements of Petrouschka). You can check them all out on this playlist. See where the Odroid N2+ really shines and where it takes a nose dive and crashes.
https://www.youtube.com/playlist?list=P...z0JrQRAfnF

Notes: MIDI files are competition/performance recordings from the Yamaha Signature MIDI Collection for their Disklavier series.
https://www.yamahaden.com/midi-files

I still feel paranoid that I'm doing something wrong. Little voice in the back of my head keeps asking, "Should I be getting this kind of performance from a ~US$100 computer?"

SOME CONCLUSIONS

- The performance of Pianoteq on the Odroid N2+ comes mainly from the stability of the Performance Index (approx. 34), which is achieved by assigning the 4 high-performance cores to Pianoteq (taskset command). So, if a note can be played by the N2+, it will most likely be played. Without core assignment, the wild fluctuations drop the performance drastically and general performance is poor. If you take a look at some of the Musical Stress Tests, you'll see that the N2+ Performance Index is generally rock-solid.

- At max settings, the Pianoteq/N2+ excels up to around 128 polyphony. If playing in this range, the N2+ handles the notes well and is able to keep the audio load within the limits of the very stable Performance Index (approx. 34).

- the N2+ begins to fail above 128 polyphony. In Musical Stress Test #7 (La Campanella), there are several sections at the end where the polyphony soars to 190+, this completely crashes the N2+ playback. There are some other challenging MIDI files on the Stress Test Playlist that will do the same. To achieve better high-polyphony performance will require a more sophisticated chip that can give a higher Performance Index, for example a 6-core i3/i5/i7.

- Clean Playback with Minor Red Lines - if there are minor red lines, I can usually achieve clean playback by lowering the Host and Internal Sample Rates from maximum (e.g. down to 48k/48k), for persistent red lines, lowering the Audio Buffer (latency) to around 2-4 ms also helped. BUT there are still some ghosts in the machine causing oddball / random red lines, and chasing them down can be difficult. For example, starting playback and switching to the Audio Load/Performance screen can cause a red line, PulseAudio also caused repeated red lines (my startup script disablesPulseAudio when I start Pianoteq).

- Clean Playback with High Polyphony - For high polyphony, lowering the Host Sample rate and Audio Buffer (latency) were not effective, in fact, keep them at max is just fine. The most effective way to get clean playback with high polyphony, was to lower the Internal Sample Rate (e.g. down to 32k or 24k) and/or using the Auto Optimistic/Pessimistic polyphony settings, that is what worked best. The best way to handle higher polyphony is to use a CPU which has a higher Performance Index.

- The onboard sound chip is very good. I have a suspicion that it does a better job of lowering the audio load, compared to my external audio interface! Perhaps it's because the onboard sound chips runs at 32-bits vs 24-bits? Or maybe it's because the onboard sound chip skips the USB latency? I'll have to run a few more tests to confirm.

Cool! I like the metal case and I notice you enable the SPDIF port. Internally what are the RCA jacks connected to? SPDIF, HDMI or 3.5mm audio output or can you switch that?

Here is the link to enclosure STL files as promised.

https://www.thingiverse.com/thing:4799455

Thank you! Interested in seeing how you mount it all together. I'll post pics as well when my new touchscreen arrives. (My touchscreen got commandeered for service elsewhere, so I'm waiting for another one). 

For Ubuntu minimal - did you have to do anything to get the audio working or did it install immediately?

This is cool! I appreciate the built-in vents, good call on that!

I'm curious if that makes any difference (and how to measure if there's a difference). It's included in all the optimization advice I've seen, but I've also seen some recommendations to ignore it. I also wonder if it's redundant when assigning cores to Pianoteq, since it's basically doing something similar.

FWIW - during the jackd2 install, it gives real-time permissions to the audio group by adding these lines to your system.

Those lines are stored in this file:

You could create the file manually, but I usually just install jackd2 since it does it automatically.

I only needed to do an apt upgrade to get the new kernel and edit config.ini