At the begging of July I’ve started testing one of SDS (Software Defined Storage) implementation – Atlantis USX. More information about product itself I will present in next posts. And right now I want to share the result of first performance tests I’ve made.
4 hosts: HP Blade ProLiant BL460c G6. Each host have:
- 2x 6 cores CPU (24 vCPU – with Hyper-threading)
- 196GB RAM
- 1x 146GB (15k) local disk for Hypervisor
- 1x 300GB (10k) local disk for data (used as a RAW space for Atlantis)
- 1x 1GBps NIC for management and storage traffic
- 1x 1GBps NIC for VM data traffic
Atlantis recommends 10GE NIC especially for storage network – so keep in my that all the tests was done using 1GBps network.
Hypervisor – Citrix XenServer 6.5 with SP1.
For test purpose I used IOmeter tool. Here is the configuration of IOmeter:
- One manager with one disk
- Maximum Size: 1GB
- Queue Depth: 20 (Outstanding IO’s)
- Test time: 5 minutes
- Access specification – VDI type:
- Block size: 4kB
- Read/Write ratio: 20/80
- Random/Sequential: 80/20
- IO Alignment: 128kB
I wanted to verify if Atlantis USX volumes are at least so good as classic storage solution. As a reference I used EMC VNX 5500 storage system configured for RAID 5.
I created one Hyper-converged (HC) volume (200GB size).
- 1 VM on VNX and 1 VM on HC volume.
- 3 VMs on VNX and 3 VMs on HC volume (in each case VMs was placed on the same physical host).
- 8 VMs on HC volume.
In each test I measured total/read/write values for IOps count, throughput (in MBps) and average latency. I also verified, what was the latency distribution. This test shows, how many I/O operations was measured for each specified time slots (0 to 50 uS; 50 to 100 uS; 100 to 200 uS; 200 to 500 ; 0.5 to 1 mS; 1 to 2 mS; 2 to 5 mS; 5 to 10 mS; 10 to 15 mS; 15 to 20 mS; 20 to 30 mS; 30 to 50 mS; 50 to 100 mS; 100 to 200 mS; 200 to 500 mS; 0.5 to 1 S; 1 to 2 s; 2 to 4.7 s; 4.7 to 5 s; 5 to 10 s; >= 10 s)
Here are the results and graphs.
[VNX] – EMC VNX 5500 configured for RAID 5, 2TB LUN
[HC] – Atlantis USX Hyper-converged volume, 200GB NFS export
vmX – Virtual Machine – Windows 2008 R2 with XenApp 6.5
Test1 – 1 VM on VNX and 1 VM on HC volume:
Test 2 – 3 VMs on VNX and 3 VMs on HC volume (in each case VMs was placed on the same physical host)
Test 3 – 8 VMs on HC volume
Based on the results of the tests I can confirm, that this product looks very promising and very similar to classic storage approach. And as I mention at the beginning, I used 1GBps NICs, instead of 10GE with low latency (Atlantis recomended).
Of course, this is just the beginning, and more tests has to be done, to confirm it – user load tests, HA test, etc.
Note: All test was done on Atlantis USX version 2.2.2.