Home > Storage > PowerScale (Isilon) > Industry Solutions and Verticals > Analytics > Deep Learning with Dell EMC Isilon > FIO
The procedure below shows how to use FIO to benchmark NFS I/O from multiple clients concurrently.
Mount to Isilon. Each client will mount to a different Isilon IP address.
cat hosts | xargs -i -P 0 ssh root@{} mkdir -p /mnt/isilon
cat hosts | xargs -i -P 0 ssh root@{} umount /mnt/isilon
ssh root@server1.example.com mount -t nfs 10.1.1.1:/ifs \
-o rsize=524288,wsize=524288 /mnt/isilon
ssh root@server2.example.com mount -t nfs 10.1.1.2:/ifs \
-o rsize=524288,wsize=524288 /mnt/isilon
Install FIO servers.
cat hosts | xargs -i -P 0 ssh root@{} yum -y install epel-release
cat hosts | xargs -i -P 0 ssh root@{} yum -y install fio
cat hosts | xargs -i ssh root@{} fio –version
Start FIO servers.
cat hosts | xargs -i ssh root@{} pkill fio
cat hosts | xargs -i ssh root@{} fio --server --daemonize=/tmp/fio.pid
Create a FIO job file shown below, named fio1.job. Set numjobs to the number of GPUs per host. This job file performs I/O that is similar to the TensorFlow CNN benchmark. It creates 30 files per GPU and then reads them sequentially concurrently.
[global]
name=job1
directory=/mnt/isilon/tmp/fio
time_based=1
runtime=600
ramp_time=60
ioengine=libaio
numjobs=8
create_serialize=0
iodepth=32
kb_base=1000
[job1]
rw=read
nrfiles=30
size=64GiB
bs=1024KiB
direct=1
sync=0
rate_iops=83
Run the FIO job.
mkdir -p /mnt/isilon/tmp/fio
fio --client=hosts fio1.job