That's a very ambitious take on this problem. It is pretty cool (and definitely more secure) to run OpenBSD but you can probably get most of the upside by slapping OpenWRT on the consumer router you already have. A $50 WDR3600 happily handles several VPNs, custom VLANs, an IPv6 tunnel, exotic routing, an external drive, and a Samba server while doing the typical SOHO router-y stuff, like wifi.
BTW Running my own name server has solved a lot of weird slowdowns I used to experience when browsing the web or sshing. According to namebench[1], my router doesn't even crack the top three when it comes to response time so I used to have it forward queries but in practice, after it warms up, it's more reliable and delivers a smoother experience than either my ISP or Google.
Maybe it's gotten better in recent years, but last time I tried consumer hardware (Linksys), there were actual hardware problems like overheating that didn't get any better with alternative firmware. My Linksys routers (I went through three of them before giving up) would lock up with increasing frequency as they aged, under both Tomato and DDWRT. When I went to Cisco (real Cisco, not Linksys Cisco) and Ubiquiti routers, the problems went away. For the price of a Ubiquiti router (under $100), I personally find no reason to spend time with open source firmware on crappy consumer hardware. It just works and is rock solid.
It can be both, but certainly one issue is hardware quality. The Linksys WRT54G and WRT54GL (of which I've purchased 8), are cheap electronic devices that last on average 12-18 months. I did have one make it out to 3 years, but I also had one kick the can about 6 months in - so it averages out. What's worse, is the failure case is "Intermittent performance" prior to just dying.
Cisco also has made crappy hardware in the past, I've had stacks of Cisco 2960 switches with dead ports, but they make up for it on the business side by having really good RMA processes if you have support agreements. They also make phenomenal hardware as well - I had a Cisco 7206 keep running without any human involvement whatsoever for 10+ years.
I'd be willing to pay $500-$550 for a solid 4 Port GE, 1U Rack Mountable, fully open, appliance that I could drop whatever operating system (OpenBSD preferred), that had gone through some extensive HALT (Highly Accelerated Life Testing) at temperatures -55C to +85C that suggested an average of 20 year life in typical consumer (23C) environments. I bet there are 10s of thousands of people who likewise would jump at the opportunity.
Sometime I wonder why the WRT54GL still get's hyped so much. Yes, it is probably the router that started the whole Wrt projects. But seeing that this 2002 relict still sells for 50+€ (on amazon) pains me. For the same money you get a superior device (eg a tp-link tl-wr1043nd and even that one is quite old by now) that runs Wrt just fine.
My experience with TP-Link has not been that great. I have three TP-Link devices (wireless router, Gig-E switch, and wireless AP) and I've had stability issues with all three since I got them. I regularly have to reboot each device, as each one will inevitably lock up after being powered on for a few days. I've also tried open-source firmware on the router, and it ran terribly no matter which flavor I tried; my 25Mbps WAN connection was reduced to about 5Mbps across the board. It runs at the correct speed with the stock firmware.
I'm seriously considering a build like the one in the article, however I'll probably go with a dual NIC Atom based nano-ITX board instead. Not only will it be cheaper, it will be easy to repurpose as a full fledged PC if I upgrade routers again in the future. A good example is this complete machine for $250, case and all: http://www.amazon.com/dp/B008KB5YCK
They basically get a (minimal) job done. (Though, unlike the WRT54GL, the USR Robotics modems are pretty reliable).
What I don't understand, is why nobody ever built a somewhat robust version of the WRT54GL?
Just build a robust device with a Skyworks 802.11a,b,g,n,ac FEM and maybe four GigE ports. Something plain, vanilla, that you can stick on your desk for 10 years. I bet they'd sell millions of them.
We're already seeing some large router makers (slowly) adopt OpenWRT/DD-WRT. Maybe there's a chance we can push for a status quo on using open source firmware for routers - perhaps some sort of "Android" for routers, but it needs to be fully GPL so OEMs can't lock down 95% of it like they do with Android.
With regards to the name-server bit, I honestly can't re-iterate this enough. It's amazing how many "My internet is slow" remarks really come down to DNS resolving slowly. My router has WAN failover, and I've had to manually tweak the DNS it chooses many times because even when a WAN is 'up', the DNS it uses might have latency exceeding 2+seconds for hours at a time (Comcast) even though the link itself is perfectly fast (125Mbit).
Merely setting a different DNS source immediately shows off the full potential of the link. I still need to setup my own local DNS server so that I can be rid of this problem entirely. :(
A typical $50 consumer router running OpenWRT will do everything you want except traffic shaping. The cheap consumer stuff all uses slow single-core MIPS CPUs that can only do traffic shaping at tens of megabits and consequently cannot do QoS for a fast DOCSIS or fiber connection. Of course, even if you do go to the trouble of getting an Intel-powered router, you still wouldn't use OpenBSD (or anything other than Linux) for QoS.
> you still wouldn't use OpenBSD (or anything other
> than Linux) for QoS.
Is this still the case? Especially in regards to OpenBSD managing a fast residential cable connection? I thought 5.6 yanked out the slow ALTQ code and brought OBSD close to linux QoS performance, especially for residential traffic management.
From a quick skim of the manpages and a few google searches, it looks like pf is a really poor substitute for tc. Without CoDel and friends, OpenBSD can at best implement half of a good QoS system.
Your categorical rejection of openbsd+pf is based on a cursory google search and man page comparison? I have not seen any QoS comparisons of OpenBSD 5.6 versus Linux. Did you find any?
There's nothing to compare. OpenBSD just plain doesn't have any active queue management. They used to have RED in the altq module, but it's been removed and I haven't found any mention of any other form of support for any of the common AQM algorithms. If it supports anything more advanced than classifying packets into a fixed set of priority queues with fixed packet count limits, it's well-hidden. What functionality is exposed and advertised through their man pages is simply not enough to put together a fully functional QoS system, regardless of how efficient it is with CPU time. There isn't even the theoretical possibility of OpenBSD doing good QoS unless they've got a large amount of complexity hidden and misleadingly glossed over by their documentation. In this case, running a dumb algorithm arbitrarily fast can never compete with the smarter algorithms.
Try what? The pf.conf man page section for queuing describes only functionality that is effectively equivalent to HTB or possibly HFSC on Linux. That algorithm on its own fundamentally cannot keep latency low because it has no mechanism to restrict queue length without crippling bandwidth. If OpenBSD's current HTB-equivalent is much more efficient than it used to be, that's an improvement but it's still only one of several required components and cannot get the job done on its own.
I second that motion. Try it (OpenBSD with pf).
I have not played with traffic shaping for myself but as always with OpenBSD I guess i would be surprised how well it works.
set prio, set queue and set tos seems to be the "tools" to use.
Why should I expect OpenBSD to not suffer from the universal problems of overly large queues when they haven't documented any mitigation techniques? I can't even find out how to get OpenBSD to do ECN marking without the now-removed altq module.
i'm too lazy to search, but almost every cheap modem OS and even windows and OSX were susceptible to pings or some other connection completely blocking out all honest traffic.
The latter has a complete open-source OS, you can ssh in and re-flash it yourself easily, a great community, the same TCP hardware offload, etc. I have been spec'ing out a BSD+soekris board setup for years, but when the Edgerouter came on the market it was a no-brainer. The fact that it works-out-of-box with little effort (for someone experienced with networking) is a big win, and that its quite easy to re-flash it and tweak as desired sure doesn't hurt.
While I really dig the DIY-router stuff, and was about to do it myself, Ubiquiti has sure made it hard to go that route when they can supply dang good products for the same price or less.
Edit: Added bit that this isn't a "zero effort for newbs" type product. If you've never setup a router, there'll be some research in your future to setup an Edgerouter, or BSD router.
It turns out a lot of the processing is offloaded to proprietary drivers on the EdgeRouter board. It's great for running their stock firmware, but not so great running OpenBSD. Ended up going with the new PC Engines apu1d4 board (http://www.pcengines.ch/apu.htm) rather than a Soekris. I believe it's the best of the options for an OpenBSD router.
I've been using the APU board with great success at home for a good while. I upgraded from the Alix board to make full use of my 1G/1G broadband.
Doing some speed tests I've seen speeds fluctuate around 500/700. Never really reached 1Gbit on any public speed test yet but it's helluva lot better than my old alix router.
I love the fact that the APU has an open bootloader, and as far as I can remember it was cheaper than what is mentioned of soekris here. I seem to remember the whole package costing me around 100 eur.
Different strokes for different folks. The Soekris hardware has a reputation for being industrial-strength and very resistant to harsh environmental conditions. The board requires no fans (the Ubiquiti has them) and has Intel NICs. The Soekris case fits nicely in a small cabinet in the average branch office and you don't need a proper cabinet for it. Disadvantage for some but for my little office, the Soekris is perfect. The speed of the Soekris is completely sufficient for our 100 Mbps biz class cable.
I just use the same hardware (Soekris net6501) with pfSense 2.2. Still requires a little more work than the Ubiquiti, which I might try next time. But it works pretty well 'out of the box', too.
In my experience, everything from Jupiter Broadcasting [0] is extrememly top-notch and informative. I'd highly recommend all of their shows (including BSD Now!)
Yeah, I've been Jupiter listerner for some time now mainly "Linux Action Show" http://www.jupiterbroadcasting.com/show/linuxactionshow/ they have a lot of useful info. Highly recommended podcast, and if you can support them with anything you can. I'm also thinking on building my own next DIY router. I'm tired of how limited the routers you buy off the shelf are. Currently running Asus RT-N56U with Padavan F/W https://code.google.com/p/rt-n56u/ I've also always wanted to switch to BSD distro but ports are not updated as often as "Arch Linux". If there where BSD rolling distro similar to well updated Arch Linux packages, I would consider switching.
Which ports system? I can't say anything about OpenBSD ports, but aside from some stuff that gets very little love from people, most of the FreeBSD ports tree is kept bang up to date, and binary packages appear shortly thereafter. pkgsrc, OTOH, is only released quarterly, though if you want, you can sync with their CVS repo, though, y'know, CVS.
The BSDs aren't distros, though some do have what might be called distros, such as PC BSD, pfSense, &c. being distros of FreeBSD, EdgeBSD being a distro of NetBSD.
The ports system is a rolling release system for non-base software though, though the base OS isn't. The closest BSD to come to having a rolling release schedule for the base OS is OpenBSD, with its six-month release cycle. The thing is that the BSDs can't have a rolling release schedule as is found in some Linux distros because the base OS is managed separately from the ports/packages: the core OS components aren't packaged, so there's no sense in which they can 'roll'.
Personally, I'd never use an OS with a rolling release cycle on a server. Too much can go wrong.
Personally I like to get a consumer router and put OpenWRT on it. It used to be a lot harder but it's gotten a lot more simple and effective. I have a few reflashed Netgear WNDR3700s but there are probably better ones out there that are pretty cheap too.
It takes more research and work but it's more simple than having to install everything onto a clean OS install.
Does Soekris have any competition in this space? Any time I check I can never find any viable competitors. Soekris seems to have hit the "IDA Pro sweet spot," AKA unbelievable product priced just low enough to scare away any new competitors.
Actually, soekris were pretty good but they are kinda outdated. I just got an apu1d4 from pcengines[0] and OpenBSD works like a charm on it. Half the price, much more power.
You need to look a bit closer to see the difference.
APU uses cheaper realtek NIC and soekris uses Intel's.
Soekris gives Intel QuickAssist to help encryption (VPN etc.). And soekris is showing up a new model later this year: http://soekris.com/products/net6801.html
But PC-engines boards are nice for their price. I'm considering to get one at home. That or spend a bit more on soekris box/
The NIC used is also important if you need the router to do PPPoE. Most Intel ones will support jumbo frames which can let you still get 1500 byte packets to the ISP.
This is nothing like the absurdities of the audiophile community. Routing traffic just isn't as trivial as you think it is, and going with cheap NICs that are missing important features and have lower-quality drivers will produce measurable and significant differences in objective benchmarks. If you have something like a 100Mbps cable connection, you can't use just any off-the-shelf hardware and expect to solve any problems in software.
The lower-end PFSense rebranded hardware [1] seems pretty competitive to me, with the $300 box they're expecting to release at the end of June looking especially nice in comparison to the soekris boards. Alternatively, if you're willing to spend a few dollars on RAM, there are a number of barebones systems out there with Atom C2[3,5,7]58 processors that will give ten times the performance, so you could do some file serving with the box too.
Dell T20 Xeon with a Type-1 hypervisor to run the router VM, using VT-d to isolate a multi-port NIC to the router VM. The same box can run other VMs with FreeNAS, Linux, Windows, etc.
I guess I should have been explicit and stated I was looking for a comparable system that competes with the soekris. Does the dell t20 really count as competition to a soekris? A quad port intel NIC and a Dell T20 w/ Xeon is at least twice as much as this soekris. And then there is the whole issue of size:
It depends on whether one also needs a NAS with RAID disks, or a Windows desktop, or a media server, or a personal cloud. If you need any of those, then there is nearly zero marginal size or cost to add a router VM, with more hardware headroom and OS flexibility than a special-purpose router.
Mine is through Pyramid (www.pyramid.de) who resell some Lanner stuff. Don't know if they'll sell outside Europe, and can't guarantee they're interested in selling single units to new customers. (We've had some previous business with them).
That's the Mikrotik hardware, and I don't believe RouterOS is open source. The "L3", "L4", "L5" you see on the product listings are the RouterOS license levels which control available features.
You can also run OpenWRT though installing it is described as "Not straightforward."
Allow me to save all of you who follow this guide $14 and hours of headaches: don't waste your time with the internal USB port on the Soekris net6501. The little Sandisk Cruzer drives that fit inside the case are total crap. The two that I bought lasted less than a day each. I think that writing the 4GB PFsense image to them was enough to kill them. Unfortunately for me, it didn't kill them in an obvious way. In my case, strange things started happening in PFsense. DNS became half-broken, DHCP for new clients didn't function, etc. I finally realized what was wrong and threw the USB drives in the trash and bought some of these guys and the problem was solved:
I'm using an ASUS RT-AC68U with totally open source firmware, supported by asus, with timely updateds to fix security issues. It's a 802.11ac. I really recommend it. Works very well.
I just added an OpenBSD firewall in bridge mode between my Comcast router and the rest of my network. It's implemented on a Shuttle DS57U (dual-core 1.5GHz Broadwell Celeron 3205U), with 16GB Crucial DDR3L RAM and a 128GB Crucial SSD (leftover parts). Total price: $358. It's a pretty sweet box: fanless, metal chassis, dual Intel Gigabit Ethernet. The only (minor) quibble is the CPU doesn't have AES-NI.
First thing that comes to mind is having more control of which version of *BSD you want to be based on. With pfsense you're on whichever version of FreeBSD they've worked up to.
That said I roll pfsense on a thin mini-itx intel board and its great.
My only concern would be the use of a SSD for something like this. I know with PFSense and Untangle, applications like this will shred a ssd in fairly short period of time.
The only thing such a system should write in bigger amounts are logfiles, unless you keep very detailed logs/packet captures you shouldn't get close to the write limitations of a modern SSD, which is tens of GB per day. (You might want to change the fsync-behaviour though to limit write amplification, e.g. syslog syncs the logfile to disk after every log line, which might become a problem if you have very active logs?)
I'm not sure of the exact details. I've only done a couple untangle boxes (using NexGen appliances). 6-8 months ago colleagues that do a lot more of them were still seeing a much higher failure rate across all SSDs than with spinning disks in Untangle boxes. Although I see now that NexGen offers a SSD Option on their boxes now. so things might have improved.
BTW Running my own name server has solved a lot of weird slowdowns I used to experience when browsing the web or sshing. According to namebench[1], my router doesn't even crack the top three when it comes to response time so I used to have it forward queries but in practice, after it warms up, it's more reliable and delivers a smoother experience than either my ISP or Google.
[1] https://code.google.com/p/namebench/