yyaazz's tech thoughts

I usually do my Data ONTAP tests with the simulator as I don’t have NetApp hardware in my lab. Last week we got a NetApp FAS2040 system for two weeks and I decided to make some VAAI tests.

About the configuration:

• Cisco UCS B200 blade system with two X5550 sockets and 48GB of RAM
• NetApp FAS2040 with 12×300 FC 15krpm disks. For this test I’ve created an aggregate of 9 disks.
• Cisco Nexus 5010 with 4Gb FC connection to the NetApp box

I went throught the same steps with VAAI enabled and disabled configurations. At the beginning I had a Windows VM with 40GB hard disk.

• add a new Hard Disk to the VM, 100GB, thick, cluster supported (zeroed)
• clone the VM (with the added disk) within the same LUN
• clone the VM to another LUN
• Storage VMotion the VM

Let’s see the results:

1. VMDK addition

Without VAAI: 10:30

With VAAI: 9:46

From the storage controller side the CPU usage were almost the same. With VAAI the CPU usage were 100% without that were about 95-100%, no real difference. On the other hand, with VAAI there’s no load on the host side nor on the SAN. You can see the following screenshots I made in the vSphere client when the VAAI were turned off.

So the zeroed VMDK creation, which without VAAI means 100GB zero write to the disk made some latency and some traffic for the SAN. With VAAI these graphs were near zero.

2. Clone within the same LUN

Without VAAI: 19:10

With VAAI: 15:41

Looking at the graphs the VAAI operation looks exactly the same. It took 100% CPU usage on the storage side and nothing on the host side:

As the graph above shows, the storage CPU is running at 100% and basically there’s no FC operations.

Without VAAI as you see above the CPU load were about 80-90% and there were about 3000 Ops/sec on the fibre channel adapter. The performance charts of vSphere:

3. Clone between LUNs

Without VAAI: 19:03

With VAAI: 15:37

As predictable with VAAI both the host side and the SAN side has no load and the storage CPU is working at 100%.

As my configuration has only one aggregate and raid group, cloning between LUNs means the same disks, so the graphs looks the same:

The only difference is the read latency.

4. Storage VMotion

Without VAAI: 20:26

With VAAI: 17:23

However the completed time differs a little bit from the last cloning, the graphs looks exacty the same.

I think this few examples show well how vStorage API for Array Integration moves the workload from the host and SAN side to the storage CPU. I’ve made these tests via FC connection, I’ll repeat these with iSCSI.

iSCSI performance test post is here.