The pread benchmark program measures I/O performance on read and outputs the benchmark result with BG/L information(*1) to stdout. It opens same data file from all nodes for read and each node repeats pread() with random offset and fixed buffer size. The total throughput is calculated from time spent on pread(). You can download this program from here.
There are some sample results from BG/L installed at Argonne National Lab ; NFS and PVFS2. It would be helpful if you can send your pread benchmark result to us. You can reach us at email@example.com.
*1 BG/L information is generated by bginfo library. http://www-unix.mcs.anl.gov/~kazutomo/bginfo/index.html for more details.
|compile and submit|
[download and compile] $ wget http://www-unix.mcs.anl.gov/~kazutomo/pread/pread-0.2.tgz $ tar xvfz pread-0.2.tgz $ cd pread-0.2/ $ make [run] You need to create a test data first. $ dd if=/dev/urandom of=test.dat bs=1M count=1024 By default, it opens "test.dat" at the current working dir. $ cqsub -t 30 -n 128 ./pread If you want to compare two file systems, submit two jobs like $ cqsub -n 256 -t 10 ./pread -f pread-test.dat -d /pvfs/$USER $ cqsub -n 256 -t 10 ./pread -f pread-test.dat -d /nfs/$USER
SYNOPSIS pread [options] [OPTIONS] -s pread buffer size -f testfile name -d directory where testfile is located at -n number of iterations for pread() -L add extra load by stderr output -e use lseek()+read() instead of pread()