Slashdot Mirror
Virtualization Is Not All Roses
An anonymous reader writes "Vendors and magazines are all over virtualization like a rash, like it is the Saviour for IT-kind. Not always, writes analyst Andi Mann in Computerworld." I've found that when it works, it's really cool, but it does add a layer of complexity that wasn't there before. Then again, having a disk image be a 'machine' is amazingly useful sometimes.
Yes - we have quite a bit that we just put in here at my shop.
Virtualization good: Webservers, middle tier stuff, etc.
Virtualization bad: DBs, memory intensive, CPU intensive.
Biggest issue? "Surprise" systems. You might see a system and notice a "reasonable" load average, then find out once it's on a VM that it was a really horrible candidate because it has huge memory, disk, CPU, or network spikes. VMWare especially seems to hate disk spikes.
What we learned is it's the not the average as much as the high-water-marks that really matter. A system that's quiet 99.99% of the time, but spikes to 100% for 60 seconds here or there can be nasty.
An operating system should be like a light switch... simple, effective, easy to use, and designed for everyone.
I'm certified for both VMware ESX 2.5 and VMware VI3. VMware's best practices are to never use a single path, whether it be for NIC or FC HBA (storage). VMware also has Virtual Switches, which not only allows you to team NICs for load balancing and failover, but also use port groups (VLANs). You can then view pretty throughput graphs for either physical NICs or virtual adapters. It's crazy amazing(TM).
.5 seconds compared to an IBM xSeries 3850 which takes 6 minutes). Oh, and there is the whole snapshotting feature, memory and disk, which allows for point in time recovery on any host. Yea... downsides indeed.
As for "putting many workloads on a box and uptime," this writer should really take a look at VMware VI3 and Vmotion. Not only can you migrate a running VM without downtime, you can "enter maintenance mode" on a physical host, and using DRS (distributed resource scheduler) it will automatically migrate the VMs to hosts and achieve a load balance between CPU/Memory. It's crazy amazing(TM).
Lastly, just to toot a bit of the virtualization horn... VMware's HA will automatically restart your VMs on other physical hosts in your HA cluster. It's not unusual for a Win2k3 VM to boot in under 20 seconds (VMware's BIOS posts in about
Virtualization is Sysadmin Utopia. -- cvl, a Virtualization Consultant
"Modern processors are way overkill for most things they're being used for."
Right - except like I said - watch those spikes. We took a system that according to our monitoring sat at essentially 0-1% used (load average: 0.01, 0.02, 0.01) and put it on a virtual. Great idea, right?
Except for the fact that once a day it runs a report that seems fairly harmless but caused the filesystem to go Read Only due to a VMWare bug. The report lasts only about 2 minutes, but it hammers the disk in apparently just the right way.
It's the spikes you have to be careful of. Just look for your high-water-marks. If the box spikes to 90% or 100% (though the load average doesn't reflect it) it will have some issues.
An operating system should be like a light switch... simple, effective, easy to use, and designed for everyone.
We took a system that according to our monitoring sat at essentially 0-1% used (load average: 0.01, 0.02, 0.01) and put it on a virtual.
Load average is a bad way of looking at machine utilization. Load average is the average number of processes on the run queue over the last 1,5,15 minutes. Programs running exclusively I/O will be on the sleep queue while the kernel does i/o stuff, giving you a load average of near-zero even though your machine is busy scrambling for files on disk or waiting for network data. Likewise, a program that consists entirely of NOOPs will give you a load average of one (+1 per each additional instance) even if its nice value is all the way up and it is quite interruptable/is really putting zero strain on your system.
Before deciding that a machine is virtualizable, don't just look at load average. Run a real monitoring utility and look at iowait times, etc.
Reid
The Right Reverend K. Reid Wightman,