Software and hardware annotations 2011 January

This document contains only my personal opinions and calls of judgement, and where any comment is made as to the quality of anybody's work, the comment is an opinion, in my judgement.

[file this blog page at: digg del.icio.us Technorati]

110116 Sun Particularly demented VM details

While explaining the particularly demented VM setup to someone I realized that they did not get some important details, because that setup has several layers of dementedness, none of which is fatal by itself, but they all come nicely together:

• Using VMs for data or network intensive workloads is often not appropriate.
• Using virtual disks as files for the data repository can be a suboptimal choice, especially .
• Using virtual disks as files for data repositories made of a large number of small files (for example an MH/Maildir/... style mailstore) creates a bad interface between two different granularity levels and caching levels.
• Using sparse (not preallocated) virtual disks for growing large datastores of many small files can be catastrophic.
• Running the VM on a host that accesses the sparse virtual disks containing growing large datastores of many small files over NFS seems particularly demented, especially considering that one could just access the relevant datastores directly over NFS, not inside a sparse virtual disk accessed over NFS.

As to the last point, a large directory tree composed of many small files does not perform that well over NFS either, but that would be a rather better than putting it inside a sparse virtual disk accessed over NFS.

In particular because then that directory tree could be backed on the NFS server itself, saving a colossal number of network accesses and VM load. For a mailstore in particular tree-scanning operations happen rarely as part of the mailstore operation, as they are usually limite to the subtree for one user. But backups and other administrative tasks (such as statistics gathering) do tree-scanning does apply to the whole tree, and being able to do it on the NFS server avoiding any network accesses and the NFS protocol can be a big win.

110109 Sun Particularly demented VM setup

Well, I have been too busy to finish a large number of draft blog entries over the past few months but today I feel particularly frustrated and I would like to recommend against:

• Using any mail storage server based on one file per message because it does not scale, especially to cases where dozens of users keep never-shrinking email collection on the mail storage server.
• Putting such mail storage into a virtual machines sharing a host with several other virtual machines.
• Setting up the virtual machine disk as a sparse growing file set instead of a single preallocated and mostly contiguous file (virtual disks are one of the very few plausible uses of DM/LVM2 ).
• Putting the virtual machine disk on a slow NFS server with a single RAID6 set (holding the virtual disk of nearly 20 active virtual machines plus the home directories of their users).

Of course the above initially works well, as there are too few files and users to trigger thrashing, and it is quick and easy even for unwise and unskilled people to setup.

One of the possible consequences if one were to do all this is that the guy inheriting their mess will find that as the disk usage grows and the locality of reference goes down virtual disk usage thrashes triggering timeouts in the virtual machine's disk drivers, and that even backing up its filesystem to move it to a different setup triggers that.