Storage I/O benchmark

By admin
Published on May 27th, 2009
2 Comments

Introduction

We performed all the tests with the multi-threaded benchmark tool sysbench (version 0.4.8).

http://sysbench.sourceforge.net/

This tool was primarily written for MySQL benchs and is very I/O intensive. Nowadays it’s one of the most appreciated tool for general I/O benchs.

Read bench commands

sysbench –test=fileio –max-time=60 –max-requests=100000000 –file-num=1 –file-extra-flags=direct –file-fsync-freq=0 –file-total-size=128M –file-test-mode=rndrd cleanup

sysbench –test=fileio –max-time=60 –max-requests=100000000 –file-num=1 –file-extra-flags=direct –file-fsync-freq=0 –file-total-size=128M –file-test-mode=rndrd prepare

sysbench –test=fileio –max-time=60 –max-requests=100000000 –file-num=1 –file-extra-flags=direct –file-fsync-freq=0 –file-total-size=128M –file-test-mode=rndrd run

Write bench commands

sysbench –test=fileio –max-time=60 –max-requests=100000000  –file-num=1 –file-extra-flags=direct –file-fsync-freq=0  –file-total-size=128M –file-test-mode=rndwr cleanup

sysbench –test=fileio –max-time=60 –max-requests=100000000  –file-num=1 –file-extra-flags=direct –file-fsync-freq=0  –file-total-size=128M –file-test-mode=rndwr prepare

sysbench –test=fileio –max-time=60 –max-requests=100000000  –file-num=1 –file-extra-flags=direct –file-fsync-freq=0  –file-total-size=128M –file-test-mode=rndwr run

Forces in presence

The goal is to be the most objective as possible, so we choose the closest configuration on each instance provider.

Flexiscale

ubuntu amd64 8.04 LTS

Gogrid

Centos amd64 5.1

EC2

ubuntu amd64 9.04

NiftyName

ubuntu amd64 9.04

Read

Transfer MB/sec : we are twice fast as the fastest competitor !

  • 0.9
  • 88.9
  • 2.5
  • 178.3

EC2, which is the most popular offer, suffers of very poor read performances.

I/O requests per second : it matchs read transfer values

  • 54.4
  • 5693.2
  • 160.0
  • 11414.4

Write

Transfer MB/sec : EC2 is a (slighty) more serious competitor

  • 0.9
  • 0.7
  • 3.7
  • 24.1

I/O requests/sec

  • 56.4
  • 42.9
  • 239.3
  • 1541.6

Conclusion

NiftyName supremacy is undeniable. NiftyName does not forget performances against features. What is done is well done, we are trying to reach the highest quality level, it’s significant for a common production environment.


2 Comments

Comments

2 Responses to “Storage I/O benchmark”

  1. Julien Danjou Says:
    May 29th, 2009 at 11:12 am

    That’s nice, but I think this is a really bad benchmark.

    We know (and you also probably) nothing about the underlying hardware running the platform, nor we know the current load of the cloud.

    That means you maybe storing your maybe doing your I/O tests on NiftyName inside a ramdisk on a cloud with no load other than your benchmark, where the 3 others are storing things on hard drives and are (over)loaded by others people.

    I suggest you either precise things more clearly, or redo the test.

    To really make a benchmark, the hardware platform should be the same everywhere. Even if I recognize it’s probably not possible/easy to install the 3 others compared platform.

  2. sebastien Says:
    May 29th, 2009 at 11:24 am

    You’re right.

    at the moment NiftyName is not very easy to setup on your own computer or small set of computers.
    We’re working on it. It will be easy to bench on a lot of platforms flavors after that ;)

    Competitors don’t have open source solutions, that’s the point, we deplore it.

Leave a Reply