A novel way to massively improve the SQ of computer audio streaming

[rickca]

Deleted

[Dev]

5 minutes ago, austinpop said:

Sure - CL17 is one whole cycle slower. By itself, no big deal, but a cycle here, a cycle there, and pretty soon you have a Tour de France!

 

The higher the running frequencies, the more effect it can have on generated noise pattern (or SI stuff). I hardly doubt it will have anything to do with latency but who knows ?

[Dev]

2 minutes ago, rickca said:

I'm looking at the squeezelite command flags, and I really don't get what this one is saying.  Also note that the last optional specification is whether or not to use mmap (memory map) ... I wonder what that might do for sound quality.

 

By default it will try to mmap if the device used in "hw" parameter is memory mapped I/O, as in my case

 

  82200000-8220ffff : 0000:00:15.0
    82200000-8220ffff : xhci-hcd

 

card0    USB Audio card        usb1  --> 2616:0110    PS Audio PS
card0    STATUS  --> access: MMAP_INTERLEAVED format: S32_LE subformat: STD channels: 2 rate: 352800 (352800/1) period_size: 1411 buffer_size: 5644

 

[austinpop]

@rickca - check out http://marcoc1712.it/?p=458 for a good explanation of the ALSA (-a) parameters.

[austinpop]

Regarding the squeezelite skipping issue, I did some tweaking using Piero's suggestions. It seems to be behaving itself, although that may be an illusion.

1. I tuned the rtirq.conf and rtapp.conf files as suggested on the website, raising max priority to 95, 90, and reducing DECR to 2.
2. On the SE, I gave the highest IRQ priority to the USB interface
3. On the Dell server box, I gave the highest IRQ priority to the Ethernet interface
4. I added apps to the high priority applications list - this is analogous to raising the priority in Process Lasso in the Windows days:
   • on endpoint: add squeezelite to the list, alongside RoonAppliance
   • on server: added RoonServer, RAATServer and perl (for LMS)

I didn't have the time or patience to see if just these changes in isolation helped SQ, but it certainly didn't degrade!

[rickca]

10 minutes ago, austinpop said:

@rickca - check out http://marcoc1712.it/?p=458 for a good explanation of the ALSA (-a) parameters.

Wow, that looks like a huge cache of information.  Is there a TOC or index so I can see what topics are covered by these hundreds of pages of documentation?

[LTG2010]

4 hours ago, austinpop said:

Regarding the squeezelite skipping issue, I did some tweaking using Piero's suggestions. It seems to be behaving itself, although that may be an illusion.

I tweaked it to bits and had no skips for over an hour, then suddenly it wouldn't stop..... even with LMS as control, so went back to Roon, will give it another go with your suggestions.

[seeteeyou]

Low capacity latency storage enclosure with expander
https://patents.google.com/patent/US10082963B2/en

 

储存箱与其系统
https://patents.google.com/patent/CN107450847A/zh

 

---

 

While those patents about hard drives might not be all that meaningful as long as we've got stuff like RAM and Optane, I was actually led to something else that could be quite relevant to this thread

 

Supermicro® World Record Benchmarks for Low Latency and Financial Computation Announced at STAC Summit
https://www.supermicro.com/newsroom/pressreleases/2014/press141113_STAC.cfm

 

Supermicro, Red Hat and Solarflare Set World Record Performance Mark with Double-Digit Latency Improvement on Financial Applications
https://www.supermicro.com/newsroom/pressreleases/2018/press180621_Financial_Application_Improvement.cfm

 

New Generation Servers Optimized for Extreme Low-Latency Trading
https://www.supermicro.com/white_paper/white_paper_Low_Latency_Server.pdf

 

Internal SMC Use Only Qualified Low-Latency NICs
https://www.supermicro.com/support/resources/low_latency_NICs_X10.pdf

 

Supermicro Hyper-Speed Solutions with new Hyper-Turbo mode
https://www.supermicro.com/products/nfo/Hyper-Speed.cfm

 

X11DPU-XLL
https://www.supermicro.com/products/motherboard/Xeon/C620/X11DPU-XLL.cfm

 

X10DRU-XLL
https://www.supermicro.com/products/motherboard/Xeon/C600/X10DRU-XLL.cfm

 

Low Latency/Jitter Optimization User's Guide
https://www.supermicro.com/products/nfo/files/Hyper-Speed/low_lat_jitter_opt_guide.pdf

 

Analysis: Inside Supermicro's Hyper-Speed HFT Server
https://intelligenttradingtechnology.com/blog/analysis-inside-supermicros-hyper-speed-hft-server

Quote

As an example, Supermicro tested network performance processing messages with the UDP and TCP protocols, and determined that over clocking reduced latency by 31%. Moreover, jitter was reduced by 73%.

 

I checked that guide in PDF format and some items looked quite interesting to me. Then I found quite a few sources of Supermicro X10DRU-XLL for $854 or so

 

https://www.zones.com/site/product/index.html?id=103137976

https://www.itcreations.com/view_product.asp?product_id=96132

https://www.neweggbusiness.com/product/product.aspx?item=9b-1wk-0010-00263

https://www.ebay.com/itm/SUPERMICRO-MOTHERBOARD-FOR-SUPERMICRO-SUPERS-X10DRU-XLL/142954734002

 

Then there's also an eBay seller from Santa Clara, CA who's offering some low-latency systems for roughly $2,000

 

https://www.ebay.com/itm/1U-Supermicro-Server-X10DRU-XLL-SYS-1028UX-LL3-B8-2x-E5-2630-v3-64GB-DDR4-12GB-S-/142945328726

https://www.ebay.com/itm/1U-Supermicro-10-Bay-2-5-X10DRU-XLL-Server-2x-E5-2620v4-SAS-3-3108-RAID-R630-EQ/152912862290

 

First one could be reasonably priced if an attractive offer were accepted. Then the form factor of X10DRU-XLL is called "Proprietary Ultra/WIO" @ 17" by 17" so basically finding the right chassis could be tricky if we weren't getting a complete system somewhere.

 

(Of course some of us might be smart enough to get CSE-119UTS-R750P on its own so that would be pretty cool as well.)

 

Now the $2,000 question is, do we all wanna chip in to order one of those low-latency servers and then send Roy / Rajiv / Larry / Piero that as Christmas gift + for testing purpose? We'll still need to get a low-latency NIC (or two if another one were added to either Dawson Canyon or Bean Canyon NUC with M.2 adapters / Thunderbolt enclosure) for Mellanox / Solarflare. A piece of X2522 would cost 200 bucks while we're talking about over $900 for X2522-PLUS

 

http://www.colfaxdirect.com/store/pc/viewPrd.asp?idproduct=3389

http://www.colfaxdirect.com/store/pc/viewPrd.asp?idproduct=3390

 

NIC could be returned easily if that were "nothing special" for Roon / LMS / HQP but not sure about a refurbished server, though we still have to ask Roy / Rajiv / Larry / Piero before we get the ball rolling.

 

Do you guys have other suggestions?

[mourip]

7 hours ago, austinpop said:

 

 

I'm not sure I understand what your question is relative to my findings. Could you clarify?

 

 

If others who are running the Roon (or other music) servers on an AL ramroot machine, please try an experiment where the music resides in the ramroot filesystem, and tell me what you hear.

 

This was an unexpected finding for me. We knew from the Zenith products that memory buffering by squeezelite in the endpoint had a big effect on SQ. I did not expect that RAM residency of music files on the music server, upstream in the network, would also have a positive impact. This is what I'd like to know if others are hearing.

 

Of course, I still don't know how to make use of this effect, when my library is too large for the largest Optane SSD. But it is interesting, nonetheless.

 

 

 

It is most likely that your confusion over my question is due to my own lack of understanding or inability to follow the more technical turn that we have taken. I was just thinking that with a distributed music playback system like Roon where the music server just serves up the music to a separate Roon endpoint booting to a ram disk, with no internal drives and with the USB stick removed, then any music is by default always playing from ram.

[davide256]

9 hours ago, austinpop said:

 

Dev,

 

As I mentioned in my post:

 

 

So even with RAM caching on the endpoint before playback, I found the RAM residency of the music files on the server side seemed to matter. As yet, I have no good theories for this.

 

Do note, that I am not talking about RAM caching on the server side. Who knows what we would hear if the music server also implemented large memory buffers. Unlike squeezelite, where a buffering application exists, I don't know of an existing server, Roon or LMS that does this.

So maybe what we need is  2 independent processor boards where 1 plays music, the other loads RAM from media for the first machine? A TB of ram would be astronomical in cost... seems to me the lesser cost is to keep ahead of the audio processor by keeping all disk IO electrically isolated to a RAM loading board.

[ted_b]

Rajiv,

Can you describe a little more, in musicality terms, what you are hearing with RAM residency (and to a slightly lesser extent, music on the Optane)? Blacker background, better tonality or timbre, tighter etc etc.  Thx

[afrancois]

Just sent my defective NUC back to sender.

Talked to a guy on the phone about the problem and he says that changing BIOS settings is not covered by warranty. Can you believe this? Anyhow, they will have a look at the unit, but I don't have much confidence.

[lmitche]

1 hour ago, mourip said:

the more technical turn that we have taken.

Should we consider starting a good rather than evil, NUC and AL technology tuning thread leaving this for those that never made their living as UNIX/Linux SAs?

[afrancois]

13 minutes ago, lmitche said:

Should we consider a technology thread for those that never made their living as UNIX SAs?

Good proposition. I've played around with Linux about 20 years ago. I had a Linux laptop setup with WiFi before most even heard of WiFi. After that, I grew tired of all the constant fiddling and never made it past a beginner.

[ted_b]

12 minutes ago, lmitche said:

Should we consider starting a good rather than evil, NUC and AL technology tuning thread leaving this for those that never made their living as UNIX/Linux SAs?

If it risks slowing down the discoveries and innovations, just to explain $#it to us dummies, then...no!   We'll figure it out when you guys get to some sort of end game (said somewhat seriously). 

 

[rickca]

36 minutes ago, afrancois said:

Just sent my defective NUC back to sender.

Talked to a guy on the phone about the problem and he says that changing BIOS settings is not covered by warranty. Can you believe this? Anyhow, they will have a look at the unit, but I don't have much confidence.

I guess you are too honest.  I'd contact Intel customer support rather than the place where you bought the NUC.

[afrancois]

1 minute ago, rickca said:

I guess you are too honest.  I'd contact Intel customer support rather than the place where you bought the NUC.

Always been like that :-)

[Middy]

48 minutes ago, afrancois said:

Just sent my defective NUC back to sender.

Talked to a guy on the phone about the problem and he says that changing BIOS settings is not covered by warranty. Can you believe this? Anyhow, they will have a look at the unit, but I don't have much confidence.

Go to the intel support forums. They might be able to help.  Bios is normal usage not installation of hardware.

Check out your consumer rights locally.

also.

Distance selling rights ect

Credit card company for not fit for purpose ect. 

That doesnt seem right.

Good luck 

[Middy]

41 minutes ago, lmitche said:

Should we consider starting a good rather than evil, NUC and AL technology tuning thread leaving this for those that never made their living as UNIX/Linux SAs?

As i mentioned before, it does seem like we are entering the Matrix... than a music player. I am still reticent about $29 AL not the money...

An idiots ..idiot friend guide..

 

 

[austinpop]

2 hours ago, mourip said:

 

It is most likely that your confusion over my question is due to my own lack of understanding or inability to follow the more technical turn that we have taken. I was just thinking that with a distributed music playback system like Roon where the music server just serves up the music to a separate Roon endpoint booting to a ram disk, with no internal drives and with the USB stick removed, then any music is by default always playing from ram.

 

1 hour ago, davide256 said:

So maybe what we need is  2 independent processor boards where 1 plays music, the other loads RAM from media for the first machine? A TB of ram would be astronomical in cost... seems to me the lesser cost is to keep ahead of the audio processor by keeping all disk IO electrically isolated to a RAM loading board.

 

Guys,

 

As I said earlier, I am not sure that there is a practical way to exploit this finding - unless you have a tiny music library.

 

But stepping back, here is the larger context that drove me to try this. From Roy's and Larry's experiments, followed by others including mine, here are the things that seem to matter to raise SQ:

1. Endpoint/streamer/standalone player
   1. NUC hardware as an endpoint
   2. booted to AL in RAM
   3. Roon/squeezelite executables in RAMroot (don't think anyone's tried different)
   4. PSU quality (as known before)
2. network infrastructure quality (as known before)
3. Server
   1. booted to AL in RAM
   2. Server horsepower (Roy's findings)
   3. Roon executables in RAMroot, but no need for Roon DB, can be on persistent store without loss of SQ
   4. location of music files (my last finding)

So my finding that music storage location on the server matters is in the context of Roy's larger finding that the server, as a whole, matters. Testing the "server matters" finding will have to wait until more people have tried various server configurations, but at least it is actionable.

 

My "music files in server RAM sound better" finding would best be implemented as code changes in Roon and LMS to implement large buffers like squeezelite on the endpoint, but who knows if they will feel incented to try this.

[austinpop]

1 hour ago, lmitche said:

Should we consider starting a good rather than evil, NUC and AL technology tuning thread leaving this for those that never made their living as UNIX/Linux SAs?

 

I would be for that. I have certainly contributed to making this thread far too geeky of late, but I guess all the Linux command line stuff took me back to the last time (just under 2 decades ago) that I wrote a shell script in anger (it was in ksh, not bash in those days).

 

I for one promise to dial back the geekiness, and will focus my findings from now on to higher-level cause->effect reports. 

 

Still, it would be nice to have a conflict-free AL tips and tricks thread. Would someone create one, please? I have my hands full steering this one.

[Teresa]

16 minutes ago, austinpop said:

...As I said earlier, I am not sure that there is a practical way to exploit this finding - unless you have a tiny music library...

 

I'm having a very hard time following this thread, I've read all the posts and don't understand what you all are talking about.

 

That said what is mourip's and davide256's finding? I ask as I do have a tiny music library and would be willing to help if I can. I have 52 GB of DSD and 24-bit PCM music files with 424.4 GB available on my hard drive. Most of my music is in physical formats, mostly SACDs which are not on my computer. I am playing an SACD in my Yamaha Blu-ray / SACD player as I type this.

 

If there is anything I can do to help with my ton of available space let me know. ?

[hifi25nl]

About loading to RAM the songs to be played:

 

HQPlayer all versions: max 250 ms

Squeezelite: you can set at the value you want with the -b parameter: 

https://ralph-irving.github.io/squeezelite.html

MPD: you can set at the value you want with audio_buffer_size and other parameters:

https://linux.die.net/man/5/mpd.conf

Jriver: memory playback and prebuffering in Audio options

Roon: see https://kb.roonlabs.com/Troubleshooting_Audio_Dropouts and here https://community.roonlabs.com/t/roon-buffering-song-before-playback/33483/2 

 

[rickca]

14 minutes ago, austinpop said:

Would someone create one, please?

Done.

https://www.computeraudiophile.com/forums/topic/54933-audiolinux-tuning/?tab=comments#comment-900774

 

[austinpop]

2 hours ago, ted_b said:

Rajiv,

Can you describe a little more, in musicality terms, what you are hearing with RAM residency (and to a slightly lesser extent, music on the Optane)? Blacker background, better tonality or timbre, tighter etc etc.  Thx

 

Hi Ted,

 

There is an overall increase in "quality," but I would say that in the context of the recent findings with AL and RAM residency, the 3 big factors that I notice the most are:

1. dynamics
2. transient attack
3. 3-dimensional imaging.

Here are a few tracks I've been using for this last round:

• Within by Daft Punk: https://tidal.com/track/20115560
• Chopin Sonata No. 2, 1st movement, Aimi Kobayashi: https://tidal.com/track/85912027
• Veinte Anos, Buena Vista Social Club: https://tidal.com/track/22141910

I actually own all three in high-res, but the above links allow all to listen for themselves. So what am I hearing? SQ improvements from RAM residency manifest in very telling ways:

1. On Within, the first few bars of the piano are enough to hear the difference. Larry described this brilliantly in the other shadow, peanut gallery thread:
   
   It sounds to me that I am hearing the actual signal rather than the moving average of the signal. Put another way, the full roughness of the signal is being reproduced instead of a smoothed version. It happens across the spectrum.
   
   It's like you are hearing the full transient attack of the piano notes. In the "before" case, these notes sounds softer, rounder. In the "after" case, there is more texture and immediacy to these notes. 
    
2. Coupled with the above is dynamics. It's like the sensation of unbuttoning a too-tight piece of clothing. The music sounds more open, more relaxed. On the Chopin - shout out to @manisandher for introducing me to this piece in one of his MQA shootouts - the transient attack of the piano from point #1 is again evident. But additionally, listen to the crescendos in this piece. With the RAM residency improvements, these crescendos become more thunderous, while the diminuendos are far more resolved.
    
3. Finally, we come to imaging, 3-D, holographic - pick your term. Again, I defer to Larry, who's become positively eloquent in describing the improvements since he started dabbling with NUCs and AL! He described the speakers disappearing, and the singers appearing as 3-dimensional, right there in the room with you. This is the thing I hear too. In Veinte Anos, the degree to which the 2 singers "float" in space and appear as distinct 3-dimensional subjects in the soundscape is what changes as I make improvements.

I hope this helps.