Lazy distro mirrors with squid

I have a problem that I think a lot of fellow developers probably have–I have enough computers (or virtual machines!) running the same operating system version(s) that I would benefit from a local mirror of them, but I don’t have so many systems that it’s actually reasonable for me to run a full mirror, which would entail rsyncing a bunch of content daily, much of which may be packages I would never use. And using a proxy server isn’t terribly practical, because with a bunch of semi-round-robin mirrors, it’s likely that two systems would pull the same package from different mirrors. A proxy server would have no way of knowing (ahead of time) that the two documents were actually the same.

What I wanted for a long time was a “lazy” mirror — something that would appear to my systems as a full mirror, but would act more as a proxy. When a client installed a particular version of a particular package for the first time, it would go fetch them from a “real” mirror, and then cache it for a long time. Subsequent requests for the same package from my “mirror” would be served from cache. I was convinced that this was impossible to do with a proxy server. Worse, I wanted to mirror multiple repos — Fedora and CentOS and EPEL, and maybe even Ubuntu. There’s no way squid can do that.

I was wrong. squid is pretty awesome. We just pull a few tricks:

  • Instead of using squid as a traditional proxy server that listens on port 3128, use it as a reverse proxy / accelerator that listens on port 80. (This is, incidentally, what sites like Wikipedia do.)
  • Abuse Massage the refresh_pattern rules to cache RPM files (etc.) for a very long time. Normally it is an awful, awful idea for proxy servers to do interfere with the Cache-Control / Expires headers that sites serve. But in the case of a mirror, we know that any updates to a package will necessarily bump the version number in the URL. Ergo, we can pretty safely cache RPMs indefinitely.
  • Set up name-based virtual hosting with squid, so that centos-mirror.lan and fedora-mirror.lan can point to different mirrors.

Two other important steps involve setting up cache_dir reasonably (by default, at least in the packages on CentOS 6, squid will only cache data in RAM), and bumping up maximum_object_size from the default of 4MB.

Here is the relevant section of my squid.conf. (The “irrelevant” section of my squid.conf is a bunch of acl lines that I haven’t really customized and can probably be deleted.)

# Listen on port 80, not 3128
# 'accel' tells squid that it's a reverse proxy
# 'defaultsite' sets the hostname that will be used if none is provided
# 'vhost' tells squid that it'll use name-based virtual hosting. I'm not
#   sure if this is actually needed.
http_port 80 accel defaultsite=mirror.lowell.lan vhost

# Create a disk-based cache of up to 10GB in size:
# (10000 is the size in MB. 16 and 256 seem to set how many subdirectories
#  are created, and are default values.)
cache_dir ufs /var/spool/squid 10000 16 256

# Use the LFUDA cache eviction policy -- Least Frequently Used, with
#  Dynamic Aging.
# It's more important to me to keep bigger files in cache than to keep
# more, smaller files -- I am optimizing for bandwidth savings, not latency.
cache_replacement_policy heap LFUDA

# Do unholy things with refresh_pattern.
# The top two are new lines, and probably aren't everything you would ever
# want to cache -- I don't account for VM images, .deb files, etc.
# They're cached for 129600 minutes, which is 90 days.
# refresh-ims and override-expire are described in the configuration here:
# but basically, refresh-ims makes squid check with the backend server
# when someone does a conditional get, to be cautious.
# override-expire lets us override the specified expiry time. (This is
#  illegal per the RFC, but works for our specific purposes.)
# You will probably want to tune this part.
refresh_pattern -i .rpm$ 129600 100% 129600 refresh-ims override-expire
refresh_pattern -i .iso$ 129600 100% 129600 refresh-ims override-expire
refresh_pattern ^ftp:           1440    20%     10080
refresh_pattern ^gopher:        1440    0%      1440
refresh_pattern -i (/cgi-bin/|\?) 0     0%      0
refresh_pattern .               0       20%     4320

# This is OH SO IMPORTANT: squid defaults to not caching objects over
# 4MB, which may be a reasonable default, but is awful behavior on our
# pseudo-mirror. Let's make it 4GB:
maximum_object_size 4096 MB

# Now, let's set up several mirrors. These work sort of like Apache
# name-based virtual hosts -- you get different content depending on
# which hostname you use in your request, even on the same IP. This lets
# us mirror more than one distro on the same machine.

# cache_peer is used here to set an upstream origin server:
#   '' is the hostname of the mirror I connect to.
#   'parent' tells squid that that this is a 'parent' server, not a peer
#    '80 0' sets the HTTP port (80) and ICP port (0)
#    'no-query' stops ICP queries, which should only be used between squid servers
#    'originserver' tells squid that this is a server that originates content,
#      not another squid server.
#    'name=as6453' tags it with a name we use on the next line.
# cache_peer_domain is used for virtual hosting.
#    'as6453' is the name we set on the previous line (for cache_peer)
#    subsequent words are virtual hostnames it answers to. (This particular
#     mirror has Fedora and Debian content mirrored.) These are the hostnames
#     you set up and will use to access content.
# Taken together, these two lines tell squid that, when it gets a request for
#  content on fedora-mirror.lowell.lan or debian-mirror.lowell.lan, it should
#  route the request to and cache the result.
cache_peer parent 80 0 no-query originserver name=as6453
cache_peer_domain as6453 fedora-mirror.lowell.lan debian-mirror.lowell.lan

# Another, for CentOS:
cache_peer parent 80 0 no-query originserver name=harvard
cache_peer_domain harvard centos-mirror.lowell.lan

You will really want to customize this. The and mirrors happen to be geographically close to me and very fast, but that might not be true for you. Check out the CentOS mirror list and Fedora mirror list to find something close by. (And perhaps fetch a file or two with wget to check speeds.) And I’m reasonably confident that you don’t have a lowell.lan domain in your home.

If you can find one mirror that has all the distros you need, you don’t need to bother with virtual hosts.

You can edit the respective repos in /etc/yum.repos.d/ to point to the hostnames you set up. Pay attention to whether the mirror matches the URL structure the file defaults to or not.

You can just drop the hostnames in /etc/hosts if you don’t have a home DNS server, e.g.,: fedora-mirror.lowell.lan centos-mirror.lowell.lan

Make DNS fly with dnsmasq –all-servers

The other day my friend Andrew messaged me out of the blue:

By default, when dnsmasq has more than one upstream server available, it will
send queries to just one server. Setting this flag forces dnsmasq to send all
queries to all available servers. The reply from the server which answers first
will be returned to the original requester.

Rather than finding it odd to receive a pasted chunk of a manpage via IM, I was fascinated by the content. By configuring dnsmasq to be a slightly obnoxious netizen, you can effectively always have your queries answered by the fastest server — whatever it happens to be for that query.

What I wondered was how big of a difference it would actually make. Would there be a measurable difference? 5%? 20%? I decided to take namebench for a spin and find out.

The results left me so incredulous that I disabled dnsmasq’s cache entirely, figuring that it was maybe giving it an unfair advantage. (Though I don’t think it is — all my upstream servers are running caches.) Re-running with dnsmasq being forced to go upstream for every query, it still ended up 124% faster than the next-best choice. (Wat?!) It strains belief, but see for yourself:

dnsmasq running against the 5 resolvers in my /etc/resolv.conf, using the --all-servers flag, had a mean response time of 32.00 milliseconds. (Yes, that’s curious even, but the fastest was 12.2 and the slowest was 666.7, so I don’t think 32.00 holds any special value.) namebench compared it against my current nameserver, my ISP’s local nameserver, which averaged 71.68ms. (It had the same 12.2ms minimum, and the worst case was 3.5 seconds.)

In fairness, namebench found a couple of servers that are slightly faster than my ISP’s, and dnsmasq wasn’t 124% faster than them. But it’s still a huge improvement. The fastest is another of my ISP’s servers, with an average of 53.65ms. Level 3′s open resolver averaged 60.48ms, OpenDNS averaged 69.21ms, and Google’s averaged 75.82ms.

What’s interesting to me here is that, by asking multiple servers and always returning the fastest, dnsmasq with –all-servers ends up being considerably faster on average than any single server. It averaged 32.00ms, while the next best one averaged 53.65ms. And bear in mind that this is with dnsmasq’s cache disabled because I worried that it would bias results in favor of dnsmasq.

Graphs, for those who love them:

(Don’t mind the “Internal 192-1-1″ entry — it’s a broken internal DNS cache that slipped in erroneously. It shouldn’t be used much on the LAN and uses Comcast as its upstream, so the fact that it’s slightly faster than Comcast suggests that there was some room for caching here.)

Recall, too, that this is with caching disabled in dnsmasq! It had one hand tied behind its back and it still won by an incredible amount.

For what it’s worth, here’s my dnsmasq.conf:


Take note that you should not use the “cache-size=0″ or “no-negcache” entries — they hurt performance. (And very badly in the real world!) I included them only to rule out the possibility of dnsmasq being faster because it was working from local cache. Of course, you really shouldn’t blindly copy-and-paste this, and it’s a very vanilla setup.

I ran dnsmasq in the foreground with sudo dnsmasq -d --all-servers.

What I’m wrestling with is the question of whether this is such poor “netizenship” that it’s bad practice. With --all-servers set, dnsmasq will send a copy of your query to every nameserver in /etc/resolv.conf. That’s five entries for me, effectively increasing load on DNS servers five times. That’s pretty obnoxious. (Though having five entries in /etc/resolv.conf is unusually many.) But at the same time, perhaps the biggest advantage of dnsmasq isn’t its borderline-magical --all-servers flag, but that it’s a caching resolver, which helps reduce the number of DNS queries that make it out of your network. So perhaps the impact isn’t quite as bad as it might seem. And in an ideal world, you’d only hit the your ISP’s servers, or a free service like Google’s or OpenDNS’s public servers, which would answer out of cache. So it does feel a little selfish and like a tragedy of the commons, but I don’t think it’s actually overly harmful.

1 Or something like that. I don’t actually have a transcript of the conversation.