MUSINGS: Windows Server 2019 update, RSC performance issues with Aquantia 10GbE AQC107, and expensive audiophile server computers (like the Wolf Audio Alpha 3 SX)...

Hey everyone, as mentioned last time, this past week I've been updating my Server computer (Intel i7-7700K based, 32GB machine) over to Windows Server 2019 Standard. It has been a few years since updating to Server 2016 which was an upgrade from Server 2012 R2 before that and Server 2012 initially installed in 2013! I figure it was about time to back-up all the essential data and totally start fresh. Some of the hard disks are 10 years old (a couple of Western Digital Greens and Red drives) and also on my list to be replaced soon. As you've probably experienced, while Windows has improved significantly, over years of use, it can bog down with inefficiencies from old installs hanging around. While maintaining IT stuff is generally not much fun, like "spring cleaning", it's necessary...

My choice for the Server operating system is due to some work related functions (web serving) I needed for this machine - stability, speed and security are more important than application compatibility. In that regard the Windows Server family has over the years proved to be reliable. For my audio and video server needs, as a platform to run RoonServer and as a fast NAS, the speed is certainly appreciated.

As you've probably seen, over the years, I have been agnostic about operating systems. I've talked about Linux, Mac OS X, and of course Windows in these blog posts with relevant audio measurements as well. Overall, unless you need some specific server feature, I would not suggest that Windows Server has any audiophile significance. Likewise, I do not believe computer software including the OS makes any kind of sonic difference during playback to a good DAC - so long as there are no significant bugs and bit perfect, you can't go wrong.

Over the years, Microsoft I think has done a good job with OS upgrades; this is to be expected with maturity of the product. In fact, the Server 2019 install went completely smoothly. These days, it's efficient to use Rufus to create a bootable USB drive with the OS .iso file as it's much faster than booting off a disk. Since Server 2016, Microsoft has unified the Server and Windows 10 code base and as a result, we can consider this a specialized version of Windows 10. Much of the OS will be completely familiar and it'll feel like a typical Windows 10 machine if you install Server with the "Desktop Experience" option.

While Windows Server 2019 did not present any issue, I did run into a snag with hardware and drivers...

As I posted a couple years back, I've been using 10GbE (10 gigabit/s, 10GBASE-T) networking at home for the "backbone" of the system which runs from my main Server machine to the Workstation.

For some reason, my 10GbE machines were having difficulty connecting to my router for maintenance of settings. Interestingly, send and receive speeds were fine between the devices connected to the same 10GbE switch, but the moment I tried to transfer data between WiFi devices and the Server or access the router for changing configuration, the upload speed from the Server slowed to an unbearable crawl! Not good especially for an audiophile as WiFi Roon Remotes like my Android phone/tablet trying to access RoonServer became unusable. (I don't know why, but the Roon iOS app for the iPad was slower but still usable!)

After hours of trying out different static IP configurations, changing cables on the NETGEAR GS110MX switches, wondering if I had erroneously reconfigured a subnet, it all finally came down to the Aquantia AQtion AQC107 chip-based ASUS XG-C100C ethernet cards!

Basically, I needed to update the firmware on these cards to the latest 3.1.109 found here (or here). For the ASUS card, I needed to edit the updatedate.xml code as discussed here so it's recognized by the flash tool. And then install the newest Windows x64 drivers version 2.1.21.0 from here. Afterwards, the cards worked as expected.

One further wrinkle was that if I went into the ethernet card's Advanced setting in Device Manager, and increased the "Receive Buffers" (max 4096) and "Transmit Buffers" (max 8184) values, upload speed slowed down again to the point where I would run into speed issues even streaming 24/192 FLAC in Roon to my endpoints. To overcome this, I needed to turn "Recv Segment Coalescing (IP4)" to "Disabled" (by default "Enabled"):

Receive Segment Coalescing (RSC, also known as Large Receive Offload or LRO) should reduce risk of CPU saturation with large amounts of data transfer by coalescing up to 64KB blocks instead of the computer dealing with multiple individual 1000-1500 byte packets; a good feature for 10GbE speed servers.

[Addendum May 2021: I've noticed that if the transmit buffer is too large from the server, sometimes this could lead to Roon disconnection errors presumably because latency is becoming too long. A device driver transmit buffer of 2048 to 4096 seems fine for me. Also, to reduce latency, I turn down Interrupt Moderation Rate to "Low". Maximum Number of RSS Queues up to maximum "8 Queues" to distribute transfers through the multicore CPU.]

Advanced CPU-offload features like RSC are known to cause issues sometimes so if you ever run into unexplained network problems, it's worth turning these off to debug. Looking into this further, according to this Lenovo tech note: "Starting with Windows 2019, Microsoft is supporting RSC in the OS and performance issues have been seen when RSC is enabled in the OS and in the X722 driver" (the Intel X722 is a 10GbE-capable network card). It looks like this is something we should all keep in mind. My latest Windows 10 update as well seemed to benefit from turning RSC off; I wonder if this is the "new normal" to disable ethernet driver RSC going forward. BTW, one can open a command line prompt and look at Window's settings to see if RSC is enabled with:
     netsh int tcp show global

With this change, the network has been stable and fast since (easily >900MB/s large file transfers across the 10GbE network between SSDs on both ends, low CPU utilization). Hope this may be helpful for those running into issues with Windows Server 2019 and in particular for those with Aquantia AQC107 NICs.

--------------------

These days, with "end point" DAC/streamers and Raspberry Pi-like low-power devices being more common, IMO, it's good that we have evolved beyond the point of needing to have laptops, Intel NUCs, Mac Minis and MacBooks in our listening rooms as we might have done a decade back. I would recommend doing whatever is possible to keep more complex devices (like full computers) with fans or spinning drives out of the sound room.

But what about server computers for our music and other media connected to the network from other rooms in the house? Let's just spend a little more time talking about those expensive "audiophile" server computers and being reviewed in the press. Stuff like this recent Stereophile review of the Wolf Audio Systems Alpha 3 SX with trimmings asking for around US$10,000.

First, notice that at least this computer seems reasonably well endowed in terms of speed and features with an Intel i7-8700 hex-core CPU at 4.6GHz, 32GB DDR4 RAM, a 1TB system SSD for the OS/apps, and 2TB SSD for audio files (you can see the specs here). The software is Linux-based which likely can be had for free although companies like these want you to believe their optimizations have resulted in sonic improvements (unless evidence shown, not likely possible). The audio server software in the review machine is JRiver 24 or Roon. It's got a TEAC-based slot load laptop Blu-ray transport for ripping disks.

Suppose we take out much of the audiophile "woo" which I consider a form of "snake oil" anyways (show me evidence that Stillpoint/eXemplar Audio products, or that "Audience copper" make any difference), we can price out a similar system with the same Streacom case - thanks to the comment from JohnG with the link to the product. Furthermore, we can see in the picture on page 2 of the review that it's a mini-ITX motherboard they're using. Let's use prices off Amazon and QuietPC I found this week in US$ and select great branded components for use in our hypothetical build:

1. Intel i7-9700 CPU LGA 1151 300 Series, 65W, 8-cores ($370)
   - Notice I'm recommending an upgraded CPU here. Feel free to underclock this to save electricity and keep the system cooler since in reality music servers don't need much CPU power unless you're doing lots of DSP as discussed previously.

2. Streacom FC10 Alpha Fanless Chassis ($335) (here's a nice review from 2016)

   - mini-ITX/micro-ATX/ATX motherboards supported, can handle up to 95W CPU. Looks like a fine case but since it's passively cooled, make sure plenty of ventilation and reasonably cool ambient temperatures when running!

3. Gigabyte Z390 AORUS Pro WiFi, mini-ITX motherboard ($170)

   - Pretty well top of the line motherboard of this type. We'll need the heat sink riser, add $35 extra. Has WiFi, Intel gigabit ethernet, M.2 connector, Optane ready, a better quality onboard DAC (Realtek ALC1220-VB), and a TosLink S/PDIF audio out like Wolf Audio's.

4. Corsair CMW32GX4M2C3200C16 Vengeance RGB Pro 2x16GB DDR4 3200 RAM ($175)

   - This looks like what Wolf Audio might be using with funky RGB lighting! Remember this is a Linux system which typically needs less RAM than Windows. As just an audio server, 32GB is also overkill and you'd be perfectly fine with 16GB of slightly faster DDR4 3600 like these Corsair Vengeance LPX (2x8GB) sticks at $80 - absolutely no need for pimping the rig with RGB lighting especially since you won't see it inside the case. :-)

5. Streacom ST-ZF240 240W power supply ($150)

   - Streacom's own fanless power supply, >90% efficiency, meant for the FC10 chassis (no ATX motherboards). Should be more than enough power for CPU, motherboard, RAM and storage. Looks like it only has 2 SATA power outs; get a Molex --> SATA cable for $5 since we have 3 SATA devices in this build.

6. 1TB Samsung 860 EVO SSD as system drive ($170)
   - Seriously guys, 1TB for the OS & apps is overkill even if you have a huge Roon library and keep many backups of that! The $90 500GB Samsung 860 EVO would be enough.

7. 2TB Samsung 860 EVO SSD for audio data ($330)
   - This 4TB Samsung 860 QVO SSD looks like a good deal if you want an upgrade in capacity for <US$500.

8. Panasonic UJ-265 Slim 6X Blu-Ray writer SATA 12.7mm slot load drive ($120)
   - I see no indication that a TEAC branded laptop slot drive would be any better. Remember that the main job for this is simply to rip CDs. The Panasonic drive is also a disk writer. Unless you really need a Blu-Ray drive, a cheaper DVD drive like this $45 LiteOn DI-8A4Sh-01 might be fine. Oh yeah, you'll need a Slimline SATA Adaptor for <$10. These days, I don't even have a disk spinner on my server; the point of a server is NOT to spin disks in real-time. Do the ripping with a laptop or on another computer and copy the files to the server.

9. Streacom PCI Express 1X Slot Riser Adapter ($7)
   - Get this if you need to stick some PCI-E cards inside as Wolf Audio did with the "Flux Capacitor".

10. Wolf Audio Systems Flux Capacitor USB Card ($750)

   - While I don't believe sticking an OCXO clock on a USB card will make any audible difference, let's "drink some Kool-Aid" and say we want one of these as a nod to Wolf Audio...

11. Software: Linux (get Ubuntu 20.04 LTS 64-bit free), JRiver 26 ($60) for DLNA streaming, Roon lifetime license ($700).

Total price even with the "Flux Capacitor" USB card, JRiver 26, and lifetime Roon - <$3500 + tax/shipping.

Notice that I "shopped around" and found that QuietPC has the Streacom case cheaper than on Amazon. Here's the Streacom FC10 case manual. It'll take some time to put together but doesn't look too hard if you're handy with time on your hands (good pandemic isolation project?). You'd be saving more than $5000 for what I believe would be a higher quality machine, plus you'd have the latest JRiver version and a lifetime of Roon!

In these troubled economic times, heck, one could be generous and find a local IT geek, pay the person $1000 and get him/her to assemble the hardware, install the software, and personally make sure it runs well on your local network.

This is not the first time I've talked about expensive "audiophile" servers. We discussed the Melco N1ZS20/2 a few years back. While we can appreciate the extra engineering audio companies put into specialized components like speakers, amps, top-of-the-line DACs, isn't a server computer like this essentially built with off-the-shelf parts with a brand name sticker slapped on the front?

Statements like:

"In my opinion, the OS that runs all the apps is as important as the apps themselves, because how it is run and configured is absolutely critical to good sound. Our extremely low latency WolfOS is absolutely tuned for audio and has a big impact on sound quality."

How in the world is latency linked to "sound quality" for a server machine? Where's the logic? Where's the evidence? Perhaps there's some value in having "support"through the company. Suppose the hardware costs a total of $3500 as above, add $1000 for the company to put this together for us, that's still "just" $4500. If we buy this even at a price of $9000, is "support" worth the other $4500? Where's the added value? Oh yeah... Must be in those hyped Stillpoint standoffs, eXemplar Audio something-or-others, and Audience wires with supposed grounding / RF/EMI rejection / vibration management 'technologies'.

Don't forget that I consider the "Flux Capacitor" USB card a bit of voodoo as well (just like the JCAT USB Femto, and SOtM tX-USBexp). If you have something like a Raspberry Pi streamer or maybe a Roon-ready ethernet endpoint, don't bother with that "Flux Capacitor" USB card and knock $750 off the above price.

While dressing it up as some kind of special "audiophile" device might impress some and this could be of luxury value as a non-utilitarian benefit, remember there is no evidence anyway that modern bit-perfect servers or USB connections make any difference! (For evidence see here and discussions here.) I call BS on Jason Victor Serinus' "Sharpening" golden ears listening skillz... In fact, if the Wolf Audio server actually made an audible difference with the dCS Rossini DAC and EMM Labs DV2 DAC he used, this would suggest that dCS and EMM Labs make bad DACs for being so sensitive to the upstream equipment!

Finally, what to make of John Atkinson's measurements?! As you can see, he had to use the AudioQuest Dragonfly Red DAC (at least he used the best of the Dragonfly DACs) to show the minute squiggles down below -125dB and admitted that a good DAC like the Mytek Brooklyn showed no difference. Isn't that hilarious? Using a $200 headphone USB DAC attached to a $10,000 computer in order to find a J-Test difference down close to the noise level!?

Atkinson then summarizes the measurements with "I have no idea what they mean, especially as they are very small in absolute terms." Come on man. You know exactly what this means. There is absolutely no audible difference. Why so pretentious when we can all see that the emperor is full-frontal stark naked?

Honestly, as one of audiophile's "elder statesmen" who thankfully has held on to the importance of objective testing, why is there even any room here for mystical speculation that these results might suggest anything potentially audible!? Just because JVS and the manufacturer said so? Perhaps at near-retirement, it's time to channel one's inner J.G. Holt and remind audiophiles as he did:

"Audio as a hobby is dying, largely by its own hand. As far as the real world is concerned, high-end audio lost its credibility during the 1980s, when it flatly refused to submit to the kind of basic honesty controls (double-blind testing, for example) that had legitimized every other serious scientific endeavor since Pascal."

The huge difference now is that as consumers we can explore for ourselves using easily-accessible tools that would not have been available even a decade back. The Wolf Audio server is consumer computer technology built on commodity parts and understood as such by hobbyists. How do you think this hobby is supposed to grow and achieve any kind of respectability if you can't simply read those FFT graphs and call it for what it is? How do you think audiophiles are supposed to have confidence in the mainstream media if obvious truths are seemingly inexpressible?

By the way, isn't the subjective attribution of a "metallic edge" to the server while using >$100k of gear in a listening room with "slap echo" doing a disservice to Wolf Audio? Let's be honest. In all likelihood, I bet in a blind test, Mr. Serinus would not be able to differentiate this computer from his Roonlabs Nucleus+ at all. What a mess... Companies selling expensive products with no evidence for benefit. Subjective reviewers claiming to hear things that are questionable in admittedly suboptimal rooms. And objective testing without the courage to take a stand!

Regardless of whether the "mainstream" audiophile media addresses the plain truth, IMO, the "gig" is up with overpriced, snake oil ridden, "audiophile" computers like this. One might as well be reviewing USB cables if you believe otherwise bug-free software, bitperfect computers "sound different".

------------------

Big thank you to the 67 blind test respondents to the "Is high Harmonic Distortion in music audible?" test! Result submission is closed. In the weeks ahead, let's take a look under the hood at the data. :-)

Thanks to David S for suggesting that I have a listen to Christopher Parkening's In Spanish Style the other day - amazing classical guitar work, and yes, sounds fantastic on the big rig!

Stay safe and hope you're all enjoying the music...