Thread Rating:
  • 0 Vote(s) - 0 Average
  • 1
  • 2
  • 3
  • 4
  • 5
a-b-c-d performance veriances
Hi. I don't want to hijack your thread, but since this is about router performance, maybe both our questions are related? If not, this can be moved into it's own thread, but I thought it may be helpful to both of us to combine the questions.

So, we have recently upgraded our backhauls on the first couple hops to wireless gear that carries ~700 Mbit of throughput. We have 2 Pairs of radios - so we can call them A-B and C-D. 'A' is wired to our switch at the shop, and it is linked wirelessly to 'B' (and can carry ~700 Mbit). On the same tower as 'B', we have the start of the C-D link, so 'C' in on the same tower as 'B', and 'C' and is linked wirelessly across town to 'D'. So, if I try to diagram this, we have:


If we connect it like that, and test from a computer connected behind 'D' to a server at the shop connected to the switch behind 'A', then we can get ~700 Mbit of throughput end to end. However, when we connect between 'B' and 'C' (in order to feed the AP's on that tower) with a StarOS board, the end-to-end throughput drops to about 350 Mbit. All those Ethernet connections are GigaBit, and the boards are PC-Engines APU's (Version 4.3.4n, 3x GigaBit Ethernet, DualCore 1Ghz CPU, 2 GBytes RAM). Since the Radio Units themselves are bridges, we have the radios themselves 'A', 'B', 'C', and 'D' all addressed 10.0.0.x numbers. When we put the APU board in between 'B' and 'C', the APU board itself also has a 10.0.0.x number, and we've bridged Ether1 and Ether3, and Connection Tracking is off. When the APU board in in the middle, the end-to-end link performance is about 1/2 of capacity. If we replace the APU board with a cable, we get ~700 Mbit speed (so it's not co-location interference) and if we replace the APU board with a $40 Gigabit switch, we get full speed to.

So, I thought I'd test this in the shop. I have three APU boards - let's call them APU-A, APU-B and APU-C. If I setup APU-A and APU-C so that they route between the two, and if I do an iPerf test over an ethernet cable, I can get ~950 Mbit between them. If I run that test through APU-B (again bridging Ether1 and Ether3 on APU-B) then the APU-A-to-APU-B throughput drops to ~510 Mbit. If I change APU-B from bridging to routing, again the performance is about 510 Mbit. If I replace the APU-B board with a VENTANA in the middle, again the performance through a VENTANA Ethernet-Ethernet bridge is about 510 Mbit. If I replace the StarOS Bridge with a $40 switch, then the APU-A to APU-C throughput is ~950 Mbit.

So - I'm not necessarily reporting a problem exactly... I was just surprised by the result, and since there was this routing-performance thread, I thought I'd add this info into the mix in case it's relevant to rafamous's question. I haven't yet tested this in a bunch of different configurations, and I yet haven't tested this with other types of router operating systems, or tested with other versions of Star - and we only recently upgraded our Backhaul to make this apparent anyway.
I think the key thing is that the "bridging" in the kernel is simply not the same as a switch. If you need access at that location, put an APU in one of the switch ports and the speed problem will just go away.

Every hop on my backbone on these high speed bridges is with industrial gigabit switches between the gear. The local router taps in via a port on the switch.

So what gear did you put up and at what distances to get your 700Meg capacity?
OK, thanks, that's what I thought. I figured that an APU or a VENTANA should have the horsepower to be able to route or bridge traffic at Gigabit speeds, but I guess not. That's OK though, I'll move to putting switches directly between the Backbone Bridges, and then plugging the APU into that switch to route traffic that is for the local AP's. That's what we were going to try next anyway, so that'll work fine.
Hardware switches tend to have hardly any impact on performance vs. software. You will find that most 'smart' managed switches have a very fast processor, as well as hardware switch logic in place to sustain the performance needs of a GbE network.

To optimize the software bridge, ensure connection tracking is disabled - this has a large impact on performance.
Yes, we have connection tracking turned off. It seems that a dual or quad core 1 ghz processor shoukd be able to keep up, but perhaps not.
I replaced the StarOS router with another type (about $100), and it passes 942 Mbit per second now. This other router is also running a Linux core (Debain I think), and is a 500 Mhz Dual Core MIPS Processor - so it would not seem that the hardware speed (at least not the raw CPU speed) is the issue. APU's have Dual Core 1 Ghz processors, and VENTANTA have Quad Core 1 Ghz processors. Of course, perhaps other router types have some optimized hardware in some way for Ethernet routing, I don't know.
Interesting, since my head end router at the end of my network that connects to the ISP's core router is running Star-OS. It is a single board computer with dual pci express Intel gigabit Ethernet, as well as aPCI quad port 10/100 card for separating a couple other network segments. We have seen internet data rates of over 800Mbits passing through this router at times. Not it is not bridging, it is all routing, but it is quite capable of passing 800+Mbits at less than 15% cpu load. I can't say as to why the Ventana won't do the same, unless there is a bottleneck in the system somewhere. I would be curious to see what it could pass in a routed situation.

That being said, when connecting these radio bridges together, no matter how fast of a system I have there for routing, I use a switch there and connect the local router to another port of the switch to tap off the traffic that has to leave the backbone at that point. All my routers are running OLSR, and I also a couple of redundant 5GHz connections to get around the 24GHz units when there are short periods of heavy rain that brings down the 24GHz links. We slow down, but at least we stay online.

Forum Jump:

Users browsing this thread: 1 Guest(s)