Lab Journal: Kudos and Disappointments
As I mentioned two weeks ago, I'm starting a new lab at the University of Pittsburgh. My intent is to have a Mac-only lab which includes managed lab computers for instruments as well as a computational cluster. All with the same Xserve infrastructure.
This week is a short status update. I'm giving out some Kudos and Disappointments based on progress so far in setting up the servers.
Kudos
Apple's server documentation
Apple should probably make this easier to find, but they have an extensive set of server documentation including walk-throughs on:
In many cases, when I had bumps in the road, it was because I didn't follow Apple's directions exactly, not because something was wrong. (I still don't know why you need guest access to enable AFP automount-able home directories, but it works now.)
AFP548
The AFP548 web site is a great resource for troubleshooting Mac OS X server problems. I found lots of helpful guides.
Automated Node Setup
In the HPC Clusters Setup guide, the walkthrough guides you to creating a default client setup using LDAP. This works great -- turn on the cluster node and everything is automatically configured. There's no need to run Server Assistant, plug in a monitor, or anything else.
Disappointments
Shipping Materials
What I'd really like Apple to do is let me specify that I'm ordering a good-sized cluster and can then dictate exactly what I want packed in the boxes. Right now, I have a huge pile of leftover materials (power cords, software DVDs, rails, screws, mounting brackets, ...).
Instead, I should be able to say "look, I know I need the short rails for square hole mounting, skip the power cords since I'm running 208V and have my own, and skip the DVDs." I should be able to send boxes back to Apple for recycling. Rather than going through each box, I should have one envelope with a report of all of my Ethernet MAC addresses, hardware serial numbers, OS X serial numbers, etc.
In short, more flexible shipping and packaging for cluster configurations would be wonderful.
Serial Numbers
While I'm thinking of serials, the OS X Server serial number system isn't flexible enough. Right now I have 17 serial numbers -- one for each box. I'd much rather have one serial number for all 17 servers. And if I expand my cluster, Apple would give me a new serial for the whole package and invalidate the previous one.
Right now, you can do this for your entire site. So if you're running a huge cluster with 1000s of nodes, or an entire company, you probably have a site licensed serial already. If you only have a few Mac servers, you probably don't care about a few serials. Those of us in the middle (i.e. 8, 10, 30, 50... servers) have this annoyance.
Yes, it's not a huge deal, but it does go against my belief that Apple tends to design things to "just work."
Lights-Out Management (LOM)
The new Intel-based Xserves have a new system for remote Server Monitor use using LOM. Unfortunately, in my experience, it isn't well enough explained. Partly, much of the documentation was written for Xserve G5's and hasn't been re-written for the new server configurations.
But the big catch is that you need a separate IP for the LOM interface. Which means that when you're adding a server to the Server Monitor, you don't want the name or address of the server. You want the name and address of the LOM interface. If you don't supply this, all you get is "Waiting for server response" and no response ever appears.
Somehow, I think the correct solution is to have Server Monitor determine it's an Intel Xserve, realize it can't connect to LOM and pop up a warning. Maybe for Leopard.
In short, there's a lot to like so far about the server setup, but nothing is perfect.
Any kudos or critiques about Apple servers in science?
