In tuning a product that is a write-intensive database tool, we’ve found that performance on a VM can be wildly inconsistent. The hardest thing to explain is why we see a 30% performance drop and slow commits when moving from physical servers to the same-size physical server holding one VM generated form that server: literally the same system, but with a VM layer.
Shoot, 37 Signals had this performance impact:
Of course, layering in a VM lets us VMotion the process, if we weren’t using a lot of local disk (since the local disk doesn’t move with it). NAS? Bwa ha ha ha, please no, corporations with I/O response requirements use SAN. Sure, SAN is expensive, but it shares the disk nicely unless you need to use the tool to monitor the SAN — SAN congestion slows the tool you need to fix it 🙂
Layering a VM tends to be satisfying a corporate requirement of “everything on VM”.
When considering Tuning the MySQL InnoDB Engine, keep in mind that if it’s on a VM, you can’t guarantee accuracy of the numbers you’re using to tune.
Recent Comments