Ian Collier wrote:
> On 18 Aug 2011, at 15:35, "Sam Skipsey" <[log in to unmask]> wrote:
> ......
>> I think that we'd be happy with 2 squids in a load-balanced
>> configuration here, not least because the network distance for CVMFS
>> -> another squid is going to be about the same as that to RAL... that
>> said, there's something to be said for multiple levels of resiliency.
>
> Just bear in mind that you don't really need to configure the squids to
> know about each other, the CVMFS client knows how to failover - both
> between squid caches and between replicas - and is better at managing
> it than round robin at least.
>
> But having a pair locally is definitely a good idea, or you risk vastly
> increasing your network traffic,
That's a backhanded way of saying that cvmfs will go direct if the squid
proxy fails (but that this might cause problems).
> and more important slowing your jobs
> down. Thee is no significant risk of overloading the replicas.
So, whilst 2 squids is desirable, you won't in principle fail jobs if
you have one and it falls over (needs a reboot etc). It will generate
lots of network traffic - which might cause things to fail.
Another optimisation is that (at least AIUI), if you have a pair of
squids, you can get them to try the other one first.
Chris
>
> Ian
>
>
>
>
>
>> Sam
>>
>> On 18 August 2011 15:25, Ewan MacMahon <[log in to unmask]> wrote:
>>> Hi all,
>>>
>>> I've just had an interesting lesson in what happens to cvmfs
>>> when its one and only squid server dies underneath it, and
>>> now I'm thinking about getting some redundancy into the
>>> system.
>>>
>>> cvmfs itself supports multiple squidd, so it's just a matter
>>> of having multiple squid servers available. The Tier 1 has
>>> a couple, but for everyone else I was wondering whether it
>>> would make sense to handle this in a similar manner to how
>>> we deal with the Frontier squids and have cross-site failover?
>>>
>>> As far as I can see we could:
>>> - use the same basic idea,
>>> - use the exact same relationships, so everyone fails over
>>> their cvmfs squid to the same place they fail over their
>>> Frontier squid (which makes sense because in a lot of
>>> cases they're the same squid),
>>> - have everyone fail over to the Tier 1,
>>> - something else,
>>> - nothing at all, and just leave it to each site to run
>>> a pair of squids.
>>>
>>> I think I'd favour either the second or third options, but
>>> I'd be interested to know what everyone thinks, and indeed
>>> how everyone else with deployed cvmfs is handling this now.
>>>
>>> Ewan
>>>
|