vSphere on Open Compute – Part 2, Make it Hurt!

image

It’s been longer than I’d have liked to put some stress on the OCP box. For some history my home lab, thanks to some kind folks, inherited a G1 Intel OCP Platform. I’ve posted about it a few times:

In that second post I promised to make the box hurt a bit and give y’all an idea of how bad you can hurt it’s feelings… This is that post, if a bit delayed.

Note: Before we get started… This is not going to be scientific, why? Well, everyone and their mother has a benchmark standard. My thought is to load this up with a bunch of things, and see where it falls over.

Note 2: It will fall over on storage first. Why? Well, the NetApp 3050 I had planned on using had a CF card die. Any of my NetApp friends happen to have a spare? Or better yet, a spare unit *wink*?

Test Scenarios

Using Razor, provision 48 WWW VMs, and then use Apache Bench, to make them hurt.

VM Setup

  • 1 vCPU
  • 1 GB Ram
  • 8 GB HD, Thin Provisioned
  • Ubuntu 12.04
  • WWW Server: Lighthttpd

Tests

  • Provisioning speed
    Measured in total time from script creation to check in with Razor. For both 1 and 48 VMs

  • Static HTML – Future Post
    Basically, I’m going to throw up lighthttpd and a generic test page. From there I’m going to throw Apache Bench at it for 5 minutes. This will be done against all 48 VMs.

  • WordPress – Future Post
    Same concept as above. Basic lighthttpd install, simple WordPress install, and Apache Bench, 5 minutes.

Results

Provisioning Speed

Single VM

This took about 3 minutes. Sorry… didn’t have the script post a date/time for the single VM.

All 48

Total time: 12m 3s. (Not bad for a single sata disk… not bad at all).

Script

1..48 | %{ new-vm -name ubuntu-$_ -vmhost $vmhost -diskmb 8192 -memorymb 1024 -diskstorageformat thin -numcpu 1 -GuestID ubuntu64Guest -Datastore (get-datastore -name *local*) -NetworkName “VM Network” -confirm:$false}

$vmhost | get-vm | start-vm -runasync -confirm:$false

As a side note, it took ~1h to install ubuntu onto all 48 from start to finish.

Notes

First off, the box gets really really loud under load. Much like any Pizza box server would. Interestingly, the fans moved enough air fast enough to move the door on the closet where the server is setup.

Next up, for those of you paying attention, the hardware was a dual hex core, with 48gb ram. So, I’ve over provisioned CPU, but not memory… why not memory? Well, the single Sata disk driving this box would not have appreciated that.

Also of note on this one, is that I tried to run the test again in quick succession, and the network saturation caused all manner of stupid in the home lab. It took ~2 days before the lab forgave me..

Resources

In getting this test to run, I drew from the following places: