Overview
This page is will list
- a set of features whose performance impacts we think are worth studying
- instructions to reproduce reliable measurements
- Results
Once we have these measurements, we can apply what we've learned to make recommendations (at KvmWorkloads) based on specific workloads.
Features to measure
- drive format (qcow2, raw (file), raw (LVM), qed)
- driver interface (virtio, scsi, ide)
- cache performance (cache=none, writeback, unsafe, writethrough)
- nic model (rtl8139, virtio)
- network type (tap, user)
- swap or no swap?
- overcommit types
- none, ksm, balloon, both ksm and balloon
- for ksm, try 1 oneiric VM, 2 oneiric VMS, and 1 oneiric and 1 lucid VM.
- Guest root filesystems (xfs, jfs, ext4, ext2)
- test each with both virtio and scsi)
- Host filesystem. (xfs, jfs, ext4)
- Guest memory (128M,256M,512M, 1024M, 2048M)
- with and without swap?
- Guest smp (1,2,4,8)
Tests could include kernbench, dbench, unixbench.
Reproduction instructions
General advice: use the same system as much as possible for each test. My plan is to have one system, re-install for each test with the same preseed to ensure the same filesystem layout.