Comment by softfalcon
1 day ago
I have heard others say the same as you about Ubiquiti devices. I genuinely curious what bottlenecks you've hit.
I've only been using Ubiquiti as a pro-sumer, but it has held up well for my use case of Plex and little game servers.
I use a Synology NAS for my storage though, which is a slightly beefier mobile AMD chipset.
I'd be very interested to know what I should and shouldn't expect from my ARM based network stack though!
> I genuinely curious what bottlenecks you've hit.
1. My UDM Pro absolutely chokes and stalls with intrusion detection enabled on the firewall and 8 cameras connected. Network goes down, cameras disconnect, devices disconnect from Wi-Fi every time a car drives past a camera due to AI features triggering, etc.
For something meant for small businesses I wish they would just shove an Intel i5 or something in it. They make great switches, great APs, great everything else, just too stingy on processors on the few pieces of central equipment that people would actually be willing to spend more on.
And for a $3999 enterprise NAS with dual 25 Gbps SFP ports and 16 drives? It could surely use something more beefy than a Neoverse N2. I'd say an i7 or even i9 is warranted here.
3. The UNAS 8 I don't own but I believe it would struggle with >1Gbps links and encryption enabled
I have to agree. I only have a consumer UDM (four years old) and it's on its last legs. From day one it was using 90%+ of its RAM and hit the CPU ceiling during large file transfers. Successive updates have pushed it well beyond its limits. I have had to disable many features like VPN and IDS/IPS. I was considering upgrading to the newer Dream Router 7 but the processor is not much of an upgrade, and it only has 3GB of RAM vs my current 2GB. I don't have space for a Pro and I'm not even sure I want one. I already have an Unraid server running with more than enough compute and RAM, and I'm going to try using OPNsense. I would prefer dedicated hardware but for the cost, it's just not worth it.
1. Same here - but it's only become a problem as protect has gained features (# of cameras stayed the same). I got a UNVR Instant and all the issues went away (I have been waiting for an updated 1U NVR but still not out yet). It sucks, but otoh protect is light years better than it had been.
I dont mind using ARM for NAS, but (to be fair I have not looked in a while) the issue is they tend to not have many pcie lanes. Looks like the N2 can have up to 64 @pcie5 so if it's built well, I don't think the CPU will be too much of a bottleneck.
Hell I'll put it out there - some company should make a NAS-specific ARM chip line to make lines of way less expensive (well pre the current troubles) base NAS enclosures with lots of NVMe etc.
Yeah mine solved once I got a UNVR as well but I would have rather paid for a better processor in the UDM Pro and not needed to buy a separate UNVR.
Unifi docs say that the AI feature run directly on the camera or via optional devices like the AI Port or AI Key. Odd that it impacts your UDM Pro and wifi.
I'm sure even if the camera runs the neural net, the detection itself triggers clips to be stored, re-encoded, indexed, etc. and the UDM Pro's processor is underpowered even for this.
It's even underpowered for streaming -- I found Protect to be extremely laggy, taking often 30+ seconds to open the camera stream when 3-4 stream receivers were connected.
Yeah . Sounds like horseshit to me to be frank.
I have a udm se, 10 g3 cams, 4k bullet+ai, door entry + cam +ai, couple of the display viewports running all day and a nano hd access point and symmetric gig with intrusion etc turned on. I also have wireguard users connecting in remotely.
No problems with performance whatsoever at this point.
Ok its not enterprisy its just a small business with 20 people but seems fine to me. I run synology servers.
2 replies →
Turn off the intrusion detection and your throughput should be significantly better.
If something has features I expect to be able to use them. They should put enough CPU to make the advertised features usable in tandem.
I echo what the others say in that it's much more important to know what portions of your traffic are going to need to be processed by the CPU than it is to know how beefy the CPU is. E.g., just to give an example of the usual investigation process:
- The EdgeRouter 12P is ancient and had a weak CPU for even the time
- However, the EdgeRouter 12P has a good selection of hardware offloads for things like routing/NAT & even a hardware switch chip. These functions will often run at (or very near) line rate without touching the CPU much, and the latency/jitter/buffer handling will often be better than when even fast CPUs handle the traffic on other products.
- Buuuut there are oddball restrictions. E.g. on the newer 2.x or 3.x software streams (i.e. for the last ~5 years) hardware offload for VLAN tagged traffic on the switch does not work, and the CPU cannot switch a full 1G of traffic without choking (it gets close, but not quite). Also the hardware switch only covers a certain range of ports, some ports can only be routed or software bridged.
- Even then, if you add a bunch of advanced firewall inspection rules it's gonna run out of CPU. Quicker if it didn't have offloads for some of the work, but still easy to make it go from a solid full gigabit WAN NAT box to 100-200 mbps depending on what you enable. This can repeat for a lot of features, like VPN and so on.
As far as host networking (i.e. a server sending data out of its NIC rather than trying to be a network switch/router/firewall between segments) usually the CPU will be a limitation for other things before it's the limitation for sending things out the NIC. And a quality NIC (which these particular ones seem to be) can make that even more true in a similar, but less extreme, way as the switching/routing hardware offloads on the EdgeRouter. E.g. ZFS can be CPU heavy with all of the parity/encryption/deduplication features you can enable and trying to do that on top of using SFTP to transfer the data to a remote host in a single encrypted stream can stress the CPU even more... but this CPU also doesn't look like a typical bargain basement ARM CPU you'd find in cheaper Ubiquiti products and would probably do fine for what it has.
Basic routing and switching - expect line speed. Don't expect analysis features to run at line speed - 30-50% penalty could be normal depending on throughput.
Stay away from IPS and complicated firewall rules which usually are done in CPU, and you should be fine. HW acceleration for those (esp. TLS decryption) is a major reason fancy firewalls are very expensive. You're better off building an IDS or picking up a smaller FortiGate or Palo Alto firewall if you really want to get serious there.
The Cloud Key Gen 2 is underpowered depending what you do with it, and it runs hot. UniFi seriously needs to refresh it. (At least it’s better than the Gen 1. The Gen 1 was disastrously bad.)
The ENAS looks like fairly nice hardware. It even has ECC RAM. Not cheap, though.
Considering how long it has been I don't think we will ever see a Cloud Key Gen 3/3+.
Ubiquiti's Cloud Gateway Max or Fiber seems to be the modern replacement since they do the job of the Cloud Key while also serving as your router and firewall.