Given that I had only five active high-speed disks, a full performance test of the VNXe wasn't really possible, but I wanted to make sure the unit wasn't obviously subpar. (See the full review, "EMC VNXe 3100: Sweet entry-level NAS and SAN.") Although I attempted to wring the most I could out of the array, I'm certain there are optimization steps that could improve the results -- and likely scalability issues that might decrease them.
To provide a basic workout, I configured two iSCSI servers on the VNXe, each bound to one of the two storage processors and attached to each of the active NICs on their respective controllers. Then I created two VMFS volumes, one on each of the two servers. The three attached vSphere hosts were then configured to use round-robin load balancing across both of their dedicated iSCSI NICs to all four of the VNXe's active interfaces, effectively eliminating any of the network interfaces as a bottleneck.
Next I deployed six virtual machines -- spread two apiece across each of the three vSphere hosts -- and attached an unformatted 10GB VMDK to each VM. These VMs were split three and three across the two iSCSI volumes, providing an even load distribution across both controllers. I then used IOmeter to run tests across all six VMs. Although I ran a lot of tests, the key one for me was a completely randomized 60 percent read and 40 percent write test using a 4KB block size across the entire disk, bombarding the same five physical 300GB 15,000-rpm SAS disks with a very difficult OLTP workload (one unlikely to be found in nature). A second test involved a much easier 2MB sequential read test to evaluate overall data throughput.
The result was that the six VMs were able to achieve a total of 862 IOPS in the OLTP test, or about 172 IOPS per active physical disk and a little over 100MBps versus the sequential read test. Overall, that's a pretty decent result for a set of five disks in a RAID5 configuration. However, both tests pushed the CPU utilization on the VNXe's storage processors beyond 70 percent. Given that I didn't have more high-speed disks to test with, I couldn't evaluate how quickly CPU bandwidth would become a limiting factor, but this is one of the significant differences between the VNXe 3100 and 3300 and something to bear in mind when choosing between the two.
You're viewing Insider content