No NFS? No RAID5? HA?

[digger4zim]

Couple quick questions....

The WiKi states (paraphrased) "No NFS. NFS is often poorly implemented...etc" Surely this doesn't exclude high end NetApp, BlueArc, EMC, etc. NFS solutions from being used with properly tuned Unix/Linux clients right?

The WiKi states (again paraphrased) "No RAID 5..throughput numbers...etc." Again this surely doesn't mean that high end RAID5 from HDS, Enginio/LSI, EMC, etc. can not be used right?

Should those statements have read "No cheap NFS. No cheap RAID5."?

As for HA, it looks like supported HA is through Red Hat cluster suite on EL4 or VCS on supported systems. Anyone using/tried VMWare or Linux-HA with Zimbra? Looking to support 100k+ mailboxes with 99.999% uptime.....without needing an army of servers.

[sampowers]

I'm using a netapp export for /opt/zimbra/store only. Other stuff probably uses mmap and so I don't think I'll be moving anything else there. It's not blown up on me yet..

[dspillett]

The problem with RAID5 is the potential poor write performance produced by every block write resulting in a at least two writes (the updated block and the updated checksum block) and probably at least one read (the other block in the strip if you are using three disks). This is particularly noticeable for systems that make amny small writes, which includes Zimbra at times.

This can be exassabated by certain journalling filesystems (which can also multiply the IO required for any given write operation).

A high-end controller will only help a little with this.

Of course for a small system it isn't an issue, but if you are considering these things then I'm guessing that your planned roll-out isn't small.

[Rich Graves]

Instead of NFS, look at iSCSI. It actually works now, has less overhead and fully supports mmap and proper locking. I would recommend investing in a SAN like LeftHand, Equallogic, Compellent, Pinnacle but you could try OpenFiler or the iSCSI target now available as a "technology preview" in RHEL 5.2/CentOS.

Don't rule out FC for cost reasons -- used 2G gear can be found with better price/performance/stability than 10G iSCSI.

RAID5 can be OK for /opt/zimbra/store, but you definitely want RAID10 for data/db/index. A modern thin-provisioned, dynamic-block-migrating SAN can eliminate the distinction and cost less than a build-your-own OpenFiler if you take thin provisioning into account. We happily use Compellents, transparently mixing RAID5 and RAID10 in the same volume.

Yes, various people have used vmware, though maybe not at the 100K user level, and Brandeis is in process of migrating 9K users to ZCS5 on RHEL5 on hand-compiled Xen 3.2. Given the speed with which a VM can live-migrate or fail-reboot, I don't see much point in RHEL or Veritas clustering. AFAIK, Zimbra clustering package doesn't do anything magical to preserve sessions in progress.