|Some random non-scientific Avahi "scaling" figures
||[Feb. 27th, 2007|10:12 pm]
Trent 'Lathiat' Lloyd (トレント)
|[||Tags|||||avahi, olpc, xo||]|
Talking to sjoerd and others on IRC, (for the benefit of the OLPC project), I decided to attempt to get some kind of an idea of the amount of traffic Avahi generates on a large network.
I booted up 80 UMLs, running 220.127.116.11, on my AMD Athlon64 X2 4200+ (O/C to 2.5GHz per core), with 2GB of ram.
Each was running with 16M ram, a base debian etch install with Avahi 0.6.16.
Interestingly with 80 VMs running my memory usage looked like this:
Mem: 2076124k total, 2012064k used, 64060k free, 18436k buffers
Swap: 996020k total, 8k used, 996012k free, 1476504k cached
I configured a 'UML Switch' with a tap device on the host attached (tun1) and told each VM to come up and use avahi-autoipd to obtain a link-local IP.
I had each VM set to advertise 3 services, via the static service advertisement files
- _activity._tcp (subtype _RSSActivity._sub._activity._tcp)
- _activity._tcp (subtype _WebActivity._sub._activity._tcp)
plus it was configured with Avahi defaults so it would announce a workstation service (the default 'ssh' service was however NOT present) and the magic services that indicate what kind of services are being announced
So I started Wireshark and IPTRAF and started booting 80 VMs, at a pace of 1 every 10 seconds, after roughly 10-15 minutes the following numbers of packets were seen on the host tun1 interface
704 UDP (56.3%)
390 ARP (21.2%)
156 OTHER (12.5%)
The ARPs are for avahi-autoipd and the UDP packets are for avahi-daemon to speak mDNS, iptraf reported
Incoming Bytes: 417,391
I then gave my local machine an IP which bumped the packet count to 712, 395 and 157.
I then started 'avahi-browse _activity._tcp', this would result in 2 services from each machine being returned, following that tidying up the packet count was at
Incoming Bytes: 496,901
Outgoing Bytes: 28,787 (30 packets according to iptraf)
Now this *really* gave me machine a heart attack, as many 'linux' processes we're eating 20% CPU as possible, and took a good 10+ seconds for my machine to start responding again, I suspect if i was running the SKAS3 patch it might be a little less harsh.
I then after cancelling that, run avahi-browse -r _activity._tcp which causes Avahi to resolve each of the services, following that run
Incoming Bytes: 570,000 packets 1384
Outgoing Bytes: 185,000 packets 227
In this case most of the services were cached and I just had to resolve each one.
I forgot to watch for traffic counts, so I re-ran the above test and iptraf claimed 165kbits/second at peak for 1 5 second interval. In this time I noticed a bunch of the service resolution queries timed out, I suspect this may have to do with it causing my machine to lock hard for a bit while it does it's magic... ;)
So that's the end of my very simple basic run of basically doing some real (rather than theoretical) tests of the number of packets seen flying around with 80 hosts on a network with Avahi with a few services, and the impact of people running a browse/resolve on a popular service type.
I'm going to try comandeer some more hardware to run some faster tests and collect some more useful data.