Morgan Tocker (mtocker) wrote,
Morgan Tocker

Bonnie++ Benchmark on EC2

I'm getting in early, I had just started building my benchmarks for my talk on Amazon EC2 at the MySQL Conference and Expo next year when I discovered exactly why they say you need to run a test more than once; results can be completely unpredictable. Take for example the Sequential I/O performance on EC2, versus my home machine:

The first 3 tests were a 36G Seagate Raptor 10k RPM, a 160G Seagate SATA2 7200RPM, and a 320G Seagate IDE 7200RPM disk, running in the same machine I had at home. The last three were Amazon EC2 images. A few observations from these results:
  • The char write test seems to max out my CPU (not on graph - see raw data), so that probably explains why it's consistent across all disks.
  • My home machine is almost dead on consistent, whereas Amazon EC2 looks more like a roller-coaster
  • The small EC2 instance seems to score better? I can't figure this one out yet, I wonder if it's because the tests on the larger instances take so much longer (given xlarge has 15G memory!)
  • From a consistency point of view, I'd rather have a single 7200RPM IDE disk (my machine), than anything EC2 can provide.

Disclosure of the test:
  • Bonnie++ defaults to creating a test of twice the size of the system memory (in order to try and lessen the skew of filesystem caches). I ran with this default, but yes, my machine had much less memory, so I'll need to work on that.
  • The operating system between all tests was CentOS 5 (a default minimal install). The
    Amazon machine images were ami-08f41161 (32-bit) and ami-31c72258 (64-bit).
  • The filesystem was ext3
  • My reference system was an AMD64 3600+ w/1G RAM (64M taken up by shared video), MSI K9VGM-V Motherboard (a cheap, basic board)
  • I have not turned off the caches on my drives, since I doubt amazon has.
  • Tests on EC2 are run from the '/mnt' directory (since the file generation is large). I've seen Paul report different results on /, but it's also possible that's just more skew.

Tags: ec2, mysql, php
  • Post a new comment


    default userpic

    Your reply will be screened

    Your IP address will be recorded 

  • 1 comment