Uncategorized

I just watched the VMworld key note. If you havent seen it you can see it here: webcast replay

This year more than 17.000 is attending VMworld, 55 people of these have actually attenende every VMworld since the beginning!

View 4.5 was finally announced. It means Windows 7 support completely new GUI for the viev manager server and other cool features.

VMware vCloud Director (vCD) was announced as well. A great product for the enterprise to help carve up their cloud for external and internal customers. More about vCD here: yellow-bricks

I really recommend watching the key note.

Transparent Page Sharing is the technology in VMware that scans and reclaims pages in memory. For instance if two virtual machines have the same content in memory VMware will share them to give more memory back to other virtual machines. You could compare it to some deduplication storage techniques.

This is a great technique and actually works really well with close to no overhead.  This is enabled pr. default and will easily reclaim 20-30% memory on older processor systems.

TPS only scans and reclaims 4KB pages. With the latest hardware geneations from Intel and AMD the MMU has moved to hardware. When using this feature VMware will use Large Pages

of 2MB instead of the 4KB prior. TPS does not scan the Large Pages because they are so big they would probably not find any two that matches.

What this means is that you wont see the same amount of memory being reclaimed and freed up for other virtual machines. When moving to the latest hardware generation expect to use 20-30% more memory pr. Physical ESX host.

VMware KB

Finally they are here. We have waited for so long. With the release of ESX 4 update 2 NFS statistics is now available for NFS datastores.

This should make storage troubleshooting a lot easier and accessible. Well done VMware.

Enhancement of the esxtop/resxtop utility— vSphere 4.0 Update 2 includes an enhancement of the performance monitoring utilities, esxtop and resxtop. The esxtop/resxtop utilities now provide visibility into the performance of NFS datastores in that they display the following statistics for NFS datastores: Reads/s, writes/s, MBreads/s, MBwrtn/s, cmds/s, GAVG/s(guest latency).

Resource pools is a way to logically abstract CPU and Memory resources to groups of virtual machines. Resource pools is a very complex thing and you need to know what the different options “Reservations”, “Limits” and “Shares” does and how they work.

A very common but (sometimes) misunderstood implementation of Resource Pools is the HIGH, NORMAL, LOW approach. You basicly create three resource pools. High (with HIGH cpu and Memory shares), Normal (With Normal cpu and memory) and Low (With Low cpu and memory shares). Then you drag and drop your virtual machines into the resource pool you want.

The CPU sheduler in VMKERNEL then uses a proportional share algorithm to give CPU and Memory to the VM’s.

This is a fine approach in SOME cases, that is if you know what this actually means. A lot of my customers are doing this to prioritize their virtual machines. They then expect a virtual machine in the HIGH resource pool to always get more CPU time than a virtual machine in the LOW resource pool. This is not the case.

Lets look at the shares to see what this implementation actually means. The shares value for High is 4, Normal 2 and Low 1:

Total shares = 7 (100%)
High 4 (57%)
Normal 2 (28%)
Low 1 (14%)

What this implementation actually does is carve up your ESX host resources or Cluster resources into three resource pools. In case of contention your high virtual machines will get 57% of the total resource, Normal 28% and Low 14%. If this is what you want great!

But if you wanted to make sure a High VM always gets more access to a CPU than a machine in Low this is not the case. Lets for instance look at the following case. You drop 2 virtual machines into the High resource pool, 4 machines into the Normal and 1 machine into the low and start a cpubusy script in each VM. This means that the 2 in high will share the 57% of the total resources. The 4 in Normal will share the 28% and the ONE in low will have the 14% to its own use. Lets look at the following screenshot from ESXTOP with this excact setup:

Look at the %USED coloumn. The fbp_low1 virtual machine gets 65% CPU usage, the 4 machines in Normal gets approx. 33% and the two in High gets 96%.

Conclusion:

If you use this resource pool design to carve up you enviroment with minumum resources to your pools you got it right. But if you use this approach to always make sure a VM in HIGH always gets more CPU than a VM in Normal or Low this is NOT the case.