Unifi AP AC square reboot themselves

We’ve had really great luck with Unifi and the access points are cheap and easy to manage. Now that there are mesh systems, perhaps this will change (the Eero is supposed to be great), but for the price, nothing is as good.

However, as we added more users and intensity, suddenly the “square” Unifi AP AC has started to crash. Ashe ” I’ve not had to turn on debugging, here are notes on things to do:

  1. Get the Cloudkey, this is a $80 mini computer that runs the Unifi controller, so you can always monitor the APs remotely. At that price, it is hard to ignore
  2. To do debugging, you can look at the Unifi logs and we see that the UAP AP AC are rebooting under heavy load as much as every 15 minutes. This doesn’t seem like it’s a problem with the round ones. I recall the square ones used different chips and never quite fit in. Beside being very hot.
  3. Make sure you lower their loads, by lowering their transmit power, we moved them to “low” which seemed to help and add more AP ACs as it seemed load dependent. The access points with say less demand seem to work fine.
  4. And yes we upgraded to the latest firmware and controller software.
  5. In looking at this, I realized that I had not configured them well, at home, where it is relatively quiet setting them up for “automatic” and very wide channels (HT40 at 2.4GHz and VHT80 at 5GHz) seems to work great, but at work, there is lots of interference, so according to the forums, it makes sense to force sectorization and go to “low” power and narrow channels (HT20 for 2.4GHz and VHT40 for 5GHz). Then you can put them closer together.

The solution seems to be to move to the “round” APs and there are couple of new choices:

  1. Gen 3 or the Unifi AP HD has arrived. Although $300, these are 4x MU-MIMO and with the latest MacBook 2016, you can get 1.4Gbps out of them. To do this, you need a pair of Ethernet cables and you have to bond them, but the work.
  2. Gen 2 or Unifi AP AC Pro is the direct replacement and is nice for low noise
  3. Unifi AC Lite is lower power and then you can get really low cost for each room kind of Wifi, not a bad choice and very cheap. This means you have to have Ethernet in each room though.
  4. Unifi AP AC LR for outdoors and big distances.

Net, net, the recommendation is to move off the square APs, they have lasted four years, so maybe that’s the life, get the new circular ones and a Cloud Key for each site.

The hunt for a X299 or X399 successor to ASUS X99-WS

We’ve had good luck with the Haswell-E Xeon 1650 v3 with the ASUS X99-WS motherboard. We get a full 16 lanes of PCIe to four slots and it is very fast. Plus it has ECC for all the memory we have added. These have been great for machine learning, for computer simulation of cars and for compiling custom Linux builds.

But now with the big processors coming out it’s time to see what the alternatives are:

Intel X299

This is the logical successor, the new HEDT (high end desktop) line seems tailor made for machine learning and training. But we need something with all those lanes as we either put in four GPUs or lots of fast SSD disks on PCI Express. PC Parts Picker has a good list of available motherboards.

AMD Ryzen

The other option is to look at the new Ryzen line or even their Epyx. This has an incredible number of cores and seems to work very well for highly threaded things and paired with the very fast nVidia GTX 1080 could really be contender particularly given the lanes.



Never say No to would you like your to return next time it is connect to your Mac

OK, I had a whole bunch of spurious network adapters stuck in my macOS Sierra network preferences, so I deleted them expected that USB discover would bring them back. Well, I then got a strange error message:

Would you like <Insert name of adapter> to return next time it is connected to your Mac?

Not understanding what this meant, I just said, NO.

Well what it means is that the next time you say plug that thing in, it will not show up on your Mac, so I don’t know when you would ever say yes to this.

The solution to this is pretty message and it involves blowing away a plist (property list). Specifically you need to start terminal and do a:

sudo rm /Library/Preferences/SystemConfiguration/NetworkInterfaces.plist

Displaying to a mount television

It used to be really complicated to mount a television on the wall, now that televisions are so small and light, it’s pretty easy you just need:

  1. Television mount. The best are tilting stands. If you have the little bit of space, this $85 unit also swivels a little which is useful if you get glare in your office. Although the $20 Amazon Basics mount will work well for most common mounting.
  2. Shelf mount. Just put this underneath the television if you have a teleconference system. The other devices can just mount behind the television to make it neater
  3. Google Chromecast Ultra. This supports 4K and is just $70 and with the Chromecast application it works with Windows, Linux, Android or Apple devices by downloading the application first. As an aside if you just want to cast audio to any speaker, this $25 Chromecast Audio works great.
  4. Apple TV. This is $150 and you should hold off if you can for the 4k versio, but it is the easiest way to get ‘no-click’ support for apple devices as the Airplay protocol is built into all Apple devices.

Adding drives 512n, 512e and 4kn

Wow there are lots of things to think about for hard disks, the main one is the warranty and performance. Paying a little more for an enterprise drive makes some sense for a consumer because you are not BackBlaze. A failure takes time to resolve and with RAID arrays too many of them and you lose data. I normally get the Seagate Enterprise Capacity drives. These are 7200 rpm with a 5 year warranty and decent reviews on Newegg. They are not super fast, but for a home NAS, that doesn’t matter that much.

However with drives today, you have a choice of sector sizes and there are some tradeoffs:

  • 512n. This means a native 512 byte sector. This was very efficient with a 40MB drive (yes those were a thing), but with a 4TB drive, it is inefficient because each sector needs its own error correction (so you are about 88% efficient vs 97% for the larger 4kn drives.
  • 4kn. This means a native 4K sector, this has two advantages, first is the efficiency noted above also. With a 100 bit ECC, a single areal disk hit is less likely to take out all your error checking in a 50 bit ECC on 512n. The main issue is compatibility. With Synology for instance you cannot mix 512n/512e drives with 4kn drives. Most modern operating systems (Synology DSM 4 and higher) support them.
  • 512e. This means 512 emulated on a real 4K sector underneath. Thus, you get compatibility and efficiency with the sector format. However particularly for write, the disk controller has to read a 4K sector modify a 512 byte subsection and write it out, so it slows things down.

The net is that if you haven’t been too worried about this, get the 512e drives. But if you are doing to have a dedicated array and then getting 4kn is the better long term choice until all

Ideal disk configurations with 10TB disks in Synology 12-disk and 8-disk arrays

Ok with 10TB hard disks and 10^16 UBER, you really can’t build raid arrays anymore, because a complete read of a 10TB disk is very likely to have an unrecoverable bit error. So what’s a person to do?

Well two strategies for a 12 disk array (like the Synology DS-1512):

  1. Install RAID-1 SSDs, this is because SSDs are 10^17 and much more reliable. If you buy a 1TB set, most of the data you will use will live on these SSDs.
  2. For the remaining 10 drives, you have a few choices focused on a RAID10 set. RAID10 seems like it is not as good as RAID6 or SHR2 because it only has one disk worth of fault tolerance, .Since the RAID-1 means mirroring however, it is much less likely to generate a UBER. As an aside, we really need something that mirrors to three disks to get two disks of fault tolerance, but that’s not a thing 🙂
  3. So with the 10 remaining drives, you are much more limited because the drives now have to match exactly. One of the nice things about SHR2 (Synology Hybrid RAID) is that you don’t need need disks of the same side.
  4. With your random disks that are less than 6TB, you can keep using SHR, the odds of failure are 1/2 that of the 10TB disks.
  5. Then for your new 10TB disks, create a RAID10 array. Note that with RAID10 you also can’t extend this disk array either, so you need to create it in place. Then when you want to expand, you take the random disks and make a neat RAID10 array when you run out of storage.
  6. As an aside, you want to get 4K native for the new 10TB and then 512e for the others assuming that SHR2 is going to be used with bunches of old disks. Since Synology requires you to have 4Kn drives separate from 512e this should work nicely

So here is one disk layout.

  1. 2 x 1TB SSDs. These are the caching drives
  2. 4 x 10TB hard disks 4Kn. These are the “new” store, set them up as RAID10, so you have really two sets of mirrors, but you stripe the data off of two drives at once. This will give you 20TB of storage.
  3. 6 x random disks 512e in SHR2. If you have a collection of 3TB, 4TB and 6TB drives, you can gradually expand this with drives up to 6TB until you get a full 6x6TB with 2 parity so effectively 24TB of data.

Then when you run out of 54TB of storage (yes, this will happen :-), you can convert the 6 slots into 10TB x 6 in RAID10 to reach a total capacity of 2x10TB = 50TB with a good chance of recovery if the 10TB fails. When the 12TB and 14TB come out, you can buy them in groups of 4x and 6x and replace them.

For an 8-disk array, it is a little trickier:

  1. The easiest thing to do is to stick with 6TB 512e drives to get 8x6TB eventually with effectively SHR2 giving you 6x6TB or 36TB
  2. If you need more, then you use the 2 x 1TB SSDs leaving 6 drives in RAID10 at 10TB given you effectly 30TB, you can see in this configuration things will actually be faster, but in this case you will have less storage than with SHR2 and smaller drives. This is because of the bit error problem on restores.