How to migrate a non-ARC site to ARC

Work in progress

Migrate a gLite site to ARC

What's what?

Arc's GridManager and GridFTP server is roughly equivalent in general function to the gLite CE. One noticeable difference is that there is no WMS. Instead meta-scheduling (i.e. job matching) is carried out by the client side tools, and input/output buffering is handled but the Grid Manager (i.e. at the CE).

There is no YAIM; configuration is all put in a single file /etc/arc.conf. Yay!

Installation

You'll need a machine, with plenty of disk space, shared over the cluster. Works just fine with SL5, if that's what you're used to.

Add the ARC repo's to the machine. On SL5, you'll also need to add the EPEL repository (for VOMS stuff). Then you just install the packages, as per the manual.

Configuration

(Read the ARC manuals; they're pretty good. Only things that are non-obvious to someone familiar with gLite is listed here). Put the cache and sessions directories on the shared space - it's simpler to set up.

In order to run both ARC and gLite in parallel, you will want to make sure that the DN -> local user mappings are the same between both systems. ARC has hooks in place for LCMAPS, but since the move to 64bit OS's, I've not been able to get this to work as of Feb 2010. (At the moment, we're using a static grid-mapfile at Glasgow).

The machine the Grid Manager runs on does much of the data movement, and thus needs it's network configured is a suitable manner - i.e. like an SE pool node, rather than a CE.

The maxjobs and maxloads parameters need tuning. maxjobs can be set lower than the full cluster, which is useful for trial purposes. Max load limits the data transfer in; at the moment at Glasgow we have

 maxload="5 2 2"   # Params: Transfers Connections Emergency.  Normal peak load is Transfers * Connections tcp/ip streams.

for 200 job slots.

RunTime Environments

A cross between setup scripts and software tags; these will typically be handled by the VO people for their own environments.