I would like to start this blog post saying that this article was actually not planned – it just happened. Some other very smart people have been doing the hard work of going through all the new features, one by one, to explain to us the differences and improvements on vSphere 4.1
I had opportunity to do some pretty interesting performance tunning exercises with vSphere 4.0 and 4.1 using Intel Nehalem 5500 CPUs with 8 physical cores HT.
Now that vSphere 4.1 RC has been released and the NDA has been lifted I can share this information. Unfortunately I can’t give details about the applications or the performance tests executed. However, I am able to say that these tests are the exact same simulation, on the same piece of hardware and they were executed on 2 VMs with 8 vCPU, 16GB RAM, in a single physical server. During the simulation there are no contentions to be seen on memory or disk IO.
The test workload had the ability to increase user simulations and achieve 100% CPU utilisation of all 8 vCPU (maximum supported by vSphere) for both VMs through increase of number of users/transactions.
There are a number of lines on my graphs below but please concentrate on the marked ones: CPU Host Utilization (%), Transaction Response Time (latency in ms) and Number of Users.
When number of users (black line) reach 2200 the Response Time start to slip (yellow line) and when more users are loaded into the system the Response Time gets even higher; eventually getting up to 100 milliseconds. You may also picture what is happening to CPU utilization (blue line) and will notice that average CPU host utilization never really goes up to 100%.
Note: Sample interval is 20s
Now, let’s have a look at vSphere 4.1 performance improvements that VMware has been announcing.
As the number of users (dark blue line) increases, the Response Time (light blue) has minimal variation, only slipping when user load is at 3000 and CPU at 100% utilization. By itself this represents an increase of 136% on the number of supported users/transactions for this specific workload. It’s is also interesting to note that vSphere 4.1 will reach more often an average of 100% CPU utilization when the system if fully loaded, and maximum average obtained with vSphere 4.0 was around 97%.
vSphere 4.1 clearly better utilizes the resources available or make more resources available thought the reduction of overhead. The impact of the new VM Wide NUMA feature must be also considered.
Take your own conclusions of these comparisons but it’s important to mention that this is not a deterministic test and the results may vary according to the type of workload, number of VMs and number of vCPUS.