Routing Tables of Death - rtod - a kernel, routing protocol, and routing daemon stress testing tool。
rtod: Routing Tables of Death.
rtod allows for the rapid injection and removal of large numbers of distinct routes per host, as a means to test the capabilities of kernels, routing protocols, and routing daemons.
Do NOT run this on a production network. DO read this documentation.
The rtod script was developed to stress test kernel routing features, daemons, and protocols as a means to measure their behavior and make them better.
rtod is as robust as I can make it - running it in more extreme modes tends to make tcp connections fail - so it has multiple means to clean up after itself automatically. You will certainly see network routing misbehave badly - and it might take 10s of minutes for the network to recover, but it will, eventually. Usually. Still, in addition to potentially disrupting your routing until many minutes after the test ends:
You may need to clean up nohup.out or kill off processes manually.
You can easily melt a processor. Or: destroy your flash if you are keeping extensive logging on it.
You might end up with other system critical processes, like network manager or odhcpd, hung or spinning madly.
You might run your kernel into the ground.
You might run out of memory and engage the oom killer on stuff you don't want killed.
You might get your whole network into a state where you have to power cycle every router simultaneously.
You might contribute to global warming.
Be careful with this tool. Lest you think I'm being overly dire - I've already done all the above to myself while testing. I'm trying to make things better by adding compute bounds, testing under loads induced by flent, better error handling, and so on.
Setup
rtod uses the unallocated fc::/8 ULA address space to do its damage.
Make utterly sure your production network refuses to forward routes injected by this script. You can easily insert a number of routes far in excess of what most daemons, cpus, and protocols can handle.
Sample conf files are provided for babeld for setup of both the routers under test and the routers that are not. I will add conf files for bird, olsr, bmx, etc as time goes by.
Test on a small scale, first. Make absolutely sure nothing escapes.
Uses: standalone
rtod's defaults are enough to mildly stress out most mesh networks. They add 256 routes for 600 seconds. This generally does no damage.
These defaults are useful for reliably repeating routing behaviors that are expected and normal, and observing metric evolution elsewhere.
rtod -r 1024
This is close to the figure being used in several production mesh networks. This begins to stress out the cpu in mips routers. The route injection itself takes many seconds for the daemon to process.
rtod -r 2048
This starts to stress out local route distribution mechanisms over wifi multicast, and the protocol itself.
Low end ARM and MIPs CPUs start getting warm. There is often a network burp while the router getting these routes sorts them out.
Routers using slow multicast start dropping off the network.
rtod -r 4096
At this level the local network daemon tends to run out of cpu on just merging in the local kernel table. Other daemons interpreting the protocol begin to struggle also.
Low end ARM and MIPs CPUs get very warm. In fact, so do low end X86_64 platforms like the pcengines apu2.
Routers using slow multicast drop off the network completely.
rtod -r 10000
This tends to completely overburden the local daemon with "interesting" results on the rest of the network. Multicast becomes a huge bottleneck. Updates lag 10s of seconds, even minutes, behind.
The injected router falls off the network, having partially announced its tables, while other daemons strain to keep up and re-announce what they got while its already become invalid. Many. Bad. Things. Happen.
Did I mention you should not run this on a production network?
In most of my tests thus far, anything above 2000 total routes begins to degrade the connectivity of a network severely. With some tuning, I've got babeld to about 5k routes but it is still barely hanging on at that level,
rtod -r 64000
Don't do this.
Uses: with other routers in the loop
Given that I have many routers, I push tests to them via pdsh. With a smaller number of local route tables, they will generate distinct routes based on their hostnames, and not cpu bottleneck on getting routes out of the kernel.
pdsh -g chips rtod -r 128 # push out 512 routes total to my 4 "c.h.i.p"s.
This is a much better test of what actually happens with multiple routers in play, in multiple places, and more complex scenarios can be easily simulated. I have one with 32 simulated routers over a network diameter of 12 hops and 3k routes, for example.
Uses: Anycast emulation
by supplying the -H option to each invocation, we can have all routers announce they have routes to the same host, and confuse them appropriately
pdsh -g chips rtod -r 256 -H mytest
Additional command line parameters
-K - the Killme command.
This is used by rtod internally to make sure it kills itself off and flushes its routes.
When used (often in desperation) at the command line, you can say:
rtod -K -t 0
To have it flush all routes immediately.
-r routes
Injects this number of routes, in the range 1-65534.
-H hostname
Specify a different hostname to use. The hostname is used to generate a md5 hash that is then turned into a usually distinct ipv6 address.
-i IFACE
You can setup an alternative interface to use that is not a dummy.
If dummy interfaces are not available, and no iface is specified, the ip6-localnet interface is used.
-p PROTO
rtod uses a separate kernel protocol table to inject and manage routes and addresses. The default is proto 50, which is unallocated by the ietf and any other daemon I'm aware of. You can use any proto you want, but, you should not - unless you want to confuse them - inject routes into the kernel tables being managed by your daemons in the first place.
This include static and boot routes as the rtod cleanup routine will wipe those out, too.
-j jitter -J JITTER
Unimplemented, but the intent is to have routes disappear and re-appear as if they were normally being lost and re-gained.
-s Source specific offset
-s offset
must be a power of two, that will inject source specific routes for a subset of existing routes.
Unimplemented, presently.
-t expires (in seconds)
The duration of the test before routes are flushed.
-m metric
Use a different metric to insert the routes than the default.
Unimplemented.
-M Random Metric Range
Inject a random metric from 0 through -M when creating the routes.
Unimplemented.
-R route_table_dump_file
Inject an exact duplicate (rapidly) of a file saved with
ip route save proto 50 > route_table_dump_file
via
ip route restore proto 50 < route_table_dump_file
de-implemented, as it has a flaw of not inserting the expires figure, and the whole point of rtod is to be able to survive abuse such as this.
from https://github.com/dtaht/rtod
No comments:
Post a Comment