I'm going through a similar process myself right now with a customer. If you're deploying anytime soon then I think your target will be 11.6.3, unless a later service pack drops soon. At this point it is looking like 11.7 will ship in Q1 2017.
I typically use ext3 on RHEL, with the noatime mount option on the db partition(s). I have also tried ext4 but I haven't done or seen any side-by-side benchmarks. Keep the AI extents logically, and if possible physically, segregated from the rest of the database. So if you lose the DB volume, you have backups and AI. If you lose the AI volume, you still have the DB; neither is catastrophic.
Have a separate volume for scratch space with a good chunk of free space; I would say at least 5x the size of your DBs. That space can be your target location for local backups (until they are shipped off-site), dumps, idxbuild temp files, etc. You may also need this space to completely rebuild the DB. Last year I had a client with database corruption and we had to rebuild the DB in a new location, preserving the old DB to be able to extract data from it. So we had old DB, dump files, new DB, idxbuild scratch files, post-load backup; the space gets chewed up pretty quickly.
I'd want to buy several spare disks to be able to swap one in immediately if there's a failure.
How large are these databases? Will 32 GB be enough RAM for effective caching? The best way to optimize I/O is to prevent it from happening in the first place, via -B/-B2/-Bt/-Bp/-omsize/-mmax etc.