h5perf, included in hdf5, is a good start. Note that the access
pattern may not match what an actual application is trying to do.
h5perf has a ton of features, though, so you can probably find a
configuration that comes close.
==rob
···
On Wed, Jan 18, 2012 at 12:51:47PM +0000, michael bane wrote:
Is there an established HDF5 benchmark suite eg for measuring file IO including for parallel file systems?
--
Rob Latham
Mathematics and Computer Science Division
Argonne National Lab, IL USA
Rob
Thanks for that. We've no parallel filesys so HDF5 build serially only
(ie on top of NFS), so have tried h5perf_serial. Just wanted to check
whether the attachments made sense, one is on the login node and other
on a compute node. The data was compiled from running:
#!/bin/bash
# script to run series of HDF5 benchmarks
# requires HDF5 module to be pre-loaded
(
for size in 20 40 60 80 100 200 400 600 800 1000; do
time ${EXE} ${RW} ${API} -e ${size},${size} -i1000
done
for size in 1000 1200 1400 1600 1800 2000; do
time ${EXE} ${RW} ${API} -e ${size},${size} -i500
done
for size in 1000 2000 3000 4000; do
time ${EXE} ${RW} ${API} -e ${size},${size} -i100
done
) 2>&1 | tee h5perf_serial_$HOSTNAME.txt
···
On 23 Jan 2012, at 16:54, Rob Latham wrote:
On Wed, Jan 18, 2012 at 12:51:47PM +0000, michael bane wrote:
> Is there an established HDF5 benchmark suite eg for measuring file
> IO including for parallel file systems?
h5perf, included in hdf5, is a good start. Note that the access
pattern may not match what an actual application is trying to do.
h5perf has a ton of features, though, so you can probably find a
configuration that comes close.
==rob
--
Rob Latham
Mathematics and Computer Science Division
Argonne National Lab, IL USA
No thoughts on that big peak? (Is it even big?!)
Thanks, M
NB:
X axis is num elements per dimension, 2 dimensions
Y axis is reported MB/s average reported
···
On 31 Jan 2012, at 23:22, michael wrote:
Rob
Thanks for that. We've no parallel filesys so HDF5 build serially only
(ie on top of NFS), so have tried h5perf_serial. Just wanted to check
whether the attachments made sense, one is on the login node and other
on a compute node. The data was compiled from running:
#!/bin/bash
# script to run series of HDF5 benchmarks
# requires HDF5 module to be pre-loaded
(
for size in 20 40 60 80 100 200 400 600 800 1000; do
time ${EXE} ${RW} ${API} -e ${size},${size} -i1000
done
for size in 1000 1200 1400 1600 1800 2000; do
time ${EXE} ${RW} ${API} -e ${size},${size} -i500
done
for size in 1000 2000 3000 4000; do
time ${EXE} ${RW} ${API} -e ${size},${size} -i100
done
) 2>&1 | tee h5perf_serial_$HOSTNAME.txt
On 23 Jan 2012, at 16:54, Rob Latham wrote:
On Wed, Jan 18, 2012 at 12:51:47PM +0000, michael bane wrote:
Is there an established HDF5 benchmark suite eg for measuring file
IO including for parallel file systems?
h5perf, included in hdf5, is a good start. Note that the access
pattern may not match what an actual application is trying to do.
h5perf has a ton of features, though, so you can probably find a
configuration that comes close.
==rob
--
Rob Latham
Mathematics and Computer Science Division
Argonne National Lab, IL USA
I've written a performance test for writing data in a chunked data set
and reading back the data in various traversal orders and with various
sizes (e.g. per line, plane).
Cheers,
Ger
"Michael Bane" 02/01/12 1:33 AM >>>
X axis is num elements per dimension, 2 dimensions
Rob
Thanks for that. We've no parallel filesys so HDF5 build serially only
(ie on top of NFS), so have tried h5perf_serial. Just wanted to check
whether the attachments made sense, one is on the login node and other
on a compute node. The data was compiled from running:
#!/bin/bash
# script to run series of HDF5 benchmarks
# requires HDF5 module to be pre-loaded
(
for size in 20 40 60 80 100 200 400 600 800 1000; do
time ${EXE} ${RW} ${API} -e ${size},${size} -i1000
done
for size in 1000 1200 1400 1600 1800 2000; do
time ${EXE} ${RW} ${API} -e ${size},${size} -i500
done
for size in 1000 2000 3000 4000; do
time ${EXE} ${RW} ${API} -e ${size},${size} -i100
done
) 2>&1 | tee h5perf_serial_$HOSTNAME.txt
On 23 Jan 2012, at 16:54, Rob Latham wrote:
On Wed, Jan 18, 2012 at 12:51:47PM +0000, michael bane wrote:
> Is there an established HDF5 benchmark suite eg for measuring file
> IO including for parallel file systems?
h5perf, included in hdf5, is a good start. Note that the access
pattern may not match what an actual application is trying to do.
h5perf has a ton of features, though, so you can probably find a
configuration that comes close.
==rob
--
Rob Latham
Mathematics and Computer Science Division
Argonne National Lab, IL USA
Rob
Thanks for that. We've no parallel filesys so HDF5 build serially only
(ie on top of NFS), so have tried h5perf_serial. Just wanted to check
whether the attachments made sense, one is on the login node and other
on a compute node. The data was compiled from running:
#!/bin/bash
# script to run series of HDF5 benchmarks
# requires HDF5 module to be pre-loaded
(
for size in 20 40 60 80 100 200 400 600 800 1000; do
time ${EXE} ${RW} ${API} -e ${size},${size} -i1000
done
for size in 1000 1200 1400 1600 1800 2000; do
time ${EXE} ${RW} ${API} -e ${size},${size} -i500
done
for size in 1000 2000 3000 4000; do
time ${EXE} ${RW} ${API} -e ${size},${size} -i100
done
) 2>&1 | tee h5perf_serial_$HOSTNAME.txt
On 23 Jan 2012, at 16:54, Rob Latham wrote:
On Wed, Jan 18, 2012 at 12:51:47PM +0000, michael bane wrote:
> Is there an established HDF5 benchmark suite eg for measuring file
> IO including for parallel file systems?
h5perf, included in hdf5, is a good start. Note that the access
pattern may not match what an actual application is trying to do.
h5perf has a ton of features, though, so you can probably find a
configuration that comes close.
==rob
--
Rob Latham
Mathematics and Computer Science Division
Argonne National Lab, IL USA