Virtualization and over-subscription: breaking the 100% utilization barrier

blog.gridcentriclabs.com

23 points·tsmith··5 comments

I can still remember the first time I put together a vmware virtualcenter server with a few servers running the hypervisor. It was the first time I encountered software that left me speechless since it felt like it was built with magic, at the time since I simply couldnt comprehend how this could be done.

I ended up quitting the company I built it at because they didnt see the point of virtualization and werent going to implement it.

Indeed. Generally speaking, virtualization adoption hasn't set any speed records. But slow and steady wins the race, as they say. :)

Also of possible interest - things we've done with the current (1.1) version of our platform:

Howto: Build and scale a Cassandra cluster in five minutes - http://blog.gridcentriclabs.com/2010/08/howto-build-and-scal...

Howto: Build a Hadoop cluster in five minutes - http://blog.gridcentriclabs.com/2010/07/howto-build-hadoop-c...

Howto: Build a ten node memcached cluster in five minutes - http://blog.gridcentriclabs.com/2010/07/how-to-build-10-core...

You may notice an interesting trend :)

[edit: formatting]

Actually OpenVZ already has a lot of the things mentioned, although only for Linux VMs running on top of an OpenVZ-ified Linux kernel.

Paging, ballooning, content-based sharing, CPU and RAM overcommit, all are there, provided you only use Linux VMs.

Good point - OpenVZ is quite good for userspace virtualization.

One thing Copper (what this demo is running on - http://www.gridcentriclabs.com/copper) can do that OpenVZ can't do is VM clone / page sharing across multiple physical servers. For example, if you wanted to clone thousands of desktops across a rack or two of servers.

Reply on news.ycombinator.com