Benchmarking the cluster

Dowload on sourceforge, then build it using (need gcc and make) :

cd src; cd current
make linux-AMD64

The execute iozone using:

./iozone [options] /directory/to/write/to/for/bench

Now to use the tool, here are the options I used :

• -i : test to use (if not specified, all of them). Tests are:

0=write/rewrite
1=read/re-read
2=random-read/write
3=Read-backwards
4=Re-write-record
5=stride-read
6=fwrite/re-fwrite
7=fread/Re-fread
8=random mix
9=pwrite/Re-pwrite
10=pread/Re-pread
11=pwritev/Re-pwritev
12=preadv/Re-preadv

• -r : block size. I use -r 8k
• -s : file size. I use -s 20G to avoid cache effects
• -t : number of threads to use. Depend of your needs. Most of the time, the number of CPU cores, to simulate an MPI/OpenMP run writing
• -a : do all the standard sizes/block combinations for this test, replace -s and -r
• -b : provide the results in an xls file to plot graphs, -b resultst.xls

Example:

./iozone -i 0 -t 5 -r 8k -s 20G /scratch/testdir

Results looks like this:

	Children see throughput for  5 initial writers 	= 1345777.16 KB/sec
	Parent sees throughput for  5 initial writers 	= 1302776.63 KB/sec
	Min throughput per process 			=  264557.00 KB/sec 
	Max throughput per process 			=  273910.25 KB/sec
	Avg throughput per process 			=  269155.43 KB/sec
	Min xfer 					= 20255392.00 KB

External links :
http://socorro.readthedocs.org/en/latest/development/diskperformancetests.html
http://www.thegeekstuff.com/2011/05/iozone-examples/