[OPEN-ILS-DEV] advice for hardware platform

[Dan Scott]

Hi Daniel:

On Wed, 2010-02-10 at 17:22 +0100, Daniël Kollmer wrote:
> Hello everyone
> 
> Well, since January we have the definitive GO! to implement Evergreen/Conifer as our ILS here at the IISG.

Awesome!

> That means I will have to design and install a hardware platform to run it on.

Oooh, fun...

> Concerning that I would like to ask for some advice of you guys who have been running it for some time.
> 
> Basically I wanted to build it up around a rather powerful dedicated server (dual quad core cpu, fast raid, 4gb memory). I have seen the schema that Laurentian uses for their setup, so I have also considered splitting off the database to run on a separate server, and the web frontend/staff client frontend too.

Your biggest bottleneck will be the database, you'll want the fast RAID
and as much memory as you can throw at the database server. If I recall
correctly, you're planning on using a separate discovery layer which
would take a lot of the read load off of the database, but cataloguers
etc will still appreciate a speedy direct connection to Evergreen. In
that case, 4GB for 8 cores probably isn't the right ratio of
memory-to-CPU; I would suggest something like 2 or 4GB RAM / CPU (the
higher, the better).

> 
> Here we are working with about 2mill records. Would that already merit a separate dedicated database machine to gain performance? Is loadbalancing of database queries something you all consider adviseable? Does it make sense to use a separate web frontend performance-wise ... and so on

You want a separate database server if you can afford it. The
Web/OpenSRF server doesn't need fast RAID or all that much RAM, so it
should be a relatively cheap piece of hardware in comparison. Say, 1 or
2GB RAM / CPU, RAID 1 for a bit of redundancy, and 8 cores.

We're not load balancing database queries in Conifer right now; we're
not even running reports against a separate database server. We just
have a main database server against which all reads and writes are
directed, plus replication going to a standby server for disaster
recovery purposes.

> Whatever tips and experiences you might like to share, I would definitely be interested in them.

I hope others weigh in!