dns masq should have resources defined

[bprashanth]

kubedns has resource/limits but not the dnsmasq cache.

[girishkalele]

100m is probably enough for CPU.
For memory resources, need to fill up dnsmasq cache with its default cache-limit using the dnsbench sample queries file and measure memory utilization.

[girishkalele]

This Google 20% project looks like another DNS benchmark datapoint we could use.

https://code.google.com/archive/p/namebench/

[bowei]

Dnsmasq memory consumption

To reproduce these results:

https://github.com/bowei/contrib/tree/dnsmasq-mem/dnsmasq-mem

TLDR

dnsmasq in its current configuration caps the size of its cache to 10k records. (See src/options.c) It will not use more than 3mb for typical workloads. When the cache limit was removed, we get the following memory usage scaling:

cache size	memory footprint
10k	~3 mb
100k	13 mb
1M	112 mb

Dnsmasq notes

• Source code has a hard limit of 10k cache entries. Performance seem to degrade when limit is increased to 1M most likely due to internal data structure choices.
• Does not cache entry types that are unbounded in memory usage (e.g.
  TXT records).
• Each record in a response consumes a single entry in the cache. E.g. if
  foo.com has 5 A records returned, then this will consume 5 entries.
• If a response with multiple records does not fit in the cache, it is not
  cached.

Evaluation details

Evaluation consists of a bind9 server configured to return synthetic records
(A, AAAA, TXT, SRV) of varying cardinality (# of result records), a dnsmasq instance and dnsperf client:

  +------+       +---------+       +---------+
  | bind | <---- | dnsmasq | <---- | dnsperf |
  +------+       +---------+       +---------+

make will build the docker images and run the memory consumption harness. The resulting run.log will contain the resulting memory used by dnsmasq after fully populating its cache.

[bprashanth]

Makes sense, so we need to define resource request on https://github.com/kubernetes/kubernetes/blob/master/cluster/addons/dns/skydns-rc.yaml.in#L85. are you suggesting 100m cpu and the default 10k/10mb memory? if cpu is left unspecified I believe it gets the lowest possible allocation under load (2 shares).

[bprashanth]

@bowei Hmm, you are a member of the Kubernetes org but I can't assign the bug to you. I think you got unassigned on my last comment, perhaps because you jumped in/out of membership somehow?

[bowei]

/close