Educate. Liberate.

So, I felt like playing around tonight, so I decided to see If I could generate some data that some might find useful. Like anything on the internet, please take this with a grain of salt… skepticism is a good thing.

I decided I would run the unixbench suite of software across the various hardwares I have available to me.

The Setup

unixbench 4.1: http://www.tux.org/pub/tux/niemi/unixbench/unixbench-4.1.0.tgz

Contestant #1
ec2-small instance
Centos 5
1.7 GB RAM
1 CPU Unit (roughly a 1.2 GHz Opeteron circa 2007, or 1 EC2 compute unit)

Contestant #2
ec2-large instance
Centos 5 64-bit
7.5 GB RAM
4 CPU Unit (2 virtual cores with 2 EC2 Compute Units each)

Contestant #3
ec2-x-large instance
Centos 5 64-bit
15 GB RAM
4 CPU Unit (4 virtual cores with 2 EC2 Compute Units each)

Contestant #4
Joyent XL accelerator
Solaris 10
2 GB RAM
1CPU Burstable up to 95%, Opteron of some flavor within 2.4-2.8GHz (based on the available offerings from sun in the X4100 series)

Contestant #5
Macbook Pro
Leopard
4 GB Ram
Core 2 duo 2.6 GHz

Unixbench was not run under complete isolation: I have no control over what other users are sucking up resources on the virtual machines and so I can’t say how outside effects where affecting the results. Nor was my laptop completely idle, although I let it sit and didn’t use it while the benchmark was running.

I can’t comment on the use of concurrency within the benchmark. It appears that every test is single threaded with the exception of the shell script tests. Take what you will from the data, but keep that in mind.

The results: http://spreadsheets.google.com/ccc?key=pYJF9YZ-X2yCZnVWs8WmgGw&hl=en

Some interesting notes

The Macbook pro cleaned up when it came to pure arithmetic, winning an overwhelming majority of those tests. Some by very large margins. I’m curious how a modern Xeon would compare.

The Joyent Accelerator showed behavior similar to my personal experience with their service: poor I/O. Some of the filesystem tests approached 50% slower than the winner, whereas the largest block size test had it off by an order of magnitude! This result could be atypical, but it seems to lend credence to my perception that I/O is inconsistent on the box.

The ec2 small instance had surprisingly poor filesystem results, falling off more sharply as the block size went down than the other machines.

The MBP had the worst filesystem performance of the big block tests for Writes.

Given these tests, the sweet spot seems to be ec2 large instance: It is very close to the XL instance in all of its marks, but it’s half the price. Its monthly price is similar to the Joyent XL, but it has 4 times the memory and significantly faster I/O