vtune amplifier options

[Pierre_B_1]

Hi,
we use vtune amplifier with IntelMPI 4.0 on nehalem architecture.
I load the module sep and pax like this :
Module Size Used by
sep3_1 56496 0
pax 22368 0

My user is in vtune group.
The hotspots collect works but the options like nehalem_general-exploration are problematic.

The process on compute node are :
gassowsk 15639 5329 0 15:49 ? 00:00:00 /opt/lsf/7.0/linux2.6-glibc2.3-x86_64/etc/res -d /opt/lsf/conf -m service2 /tmp_user/stelvio/lsf_spool/1296485356.302046
gassowsk 15641 15639 0 15:49 ? 00:00:00 /bin/sh /tmp_user/stelvio/lsf_spool/1296485356.302046
gassowsk 15645 15641 0 15:49 ? 00:00:00 /bin/sh /tmp_user/stelvio/lsf_spool/1296485356.302046.shell
gassowsk 15654 1 0 15:49 ? 00:00:00 python /tmp_user/stelvio/opt/intel/impi/3.2.1.009/bin/mpd
gassowsk 15662 15645 0 15:49 ? 00:00:00 amplxe-cl -collect nehalem_general-exploration -- mpiexec.hydra -bootstrap fork -np 8 /tmp_user/stelvio/gassowsk/TEST/gather.x
gassowsk 15671 15662 0 15:49 ? 00:00:00 /opt/intel/vtune_amplifier_xe_2011/bin64/amplxe-runsa --ui-output-format xml --ui-output-fd 4 --option-file /tmp_user/stelvio/gassowsk/TEST/r032ge/config/runsa.options
gassowsk 15707 15671 0 15:49 ? 00:00:00 /opt/intel/vtune_amplifier_xe_2011/bin64/amplxe-runsa --ui-output-format xml --ui-output-fd 4 --option-file /tmp_user/stelvio/gassowsk/TEST/r032ge/config/runsa.options

The job don't terminate and is killed with cputime limit.
Thank you for your help.

[Peter_W_Intel]

VTune Amplifier XE 2011 can only collect performance data on one node. You have to install the product on other nodes separately,and collect performance data from different node.

SoI suggest you to start data collection by using amplxe-cl but set start-paused, set duration time for all nodes. SecondaryrunMPI application without VTune Amplifier XE from one node, you can insert __itt code to do resume/paused control in MPI application (seethis article).Again, you have to set duration to terminate data collectionon all nodes.

If you want tocollect performance data of MPI application on singlenode with multicore, please refer to this article.

Regards, Peter

[Mark_D_Intel]

Here is Peter's suggestion expanded more explicitly (without the -start-paused)

In one terminal, run
> amplxe-cl -collect nehalem_general-exploration -duration 300
(or set the duration for however many seconds the run will take)

In another terminal, perform the MPI run
> mpiexec ....



If you don't know how long the run will take, there are some alternate methods for stopping collection

In one terminal, run
> amplxe-cl -collect nehalem_general-exploration -duration unlimited

In another terminal, perform the MPI run
> mpiexec ...

To stop collection, enter Ctrl-C in the first terminal, or use the 'stop' command:
> amplxe-cl -command stop

[Pierre_B_1]

Ok thanks for those suggestions.
I made a lot of tests, I re-installed the soft et the driver.
I think there was conflicts or side effects between the first installation and the method to launch amplifier
on diskless nodes (cluster SGI ICE) with login ldap, batch server LSF, MPI libraries (Intel or MPT), etc ......

On one node everythings works !
Now my problem is to find a methodology to launch amplifier-cl on multinodes and to process the results in
a uniq amplifier gui.

We would like to estimate the Flops, is it possible ?

Best regards

[Peter_W_Intel]

Sounds good - it works after uninstalling/reinstalling the product.

There is no predefined analysis type for Flops, you can define (create) yourself new analysis typeto measure Flops:
1) On left-bottom of "Choose Analysis Type" dialog, click "New..." button to select "Hardware Event-based Sampling Analysis"
2) Edit new analysis type name
3) Click "Edit" button then click "Add Event" button to add new event(s) to your new analysis type. You might select event "X87_OPS_RETIRED.ANY" or/and event "FP_ASSIST"

Regards, Peter

[Huseyin_A_Intel1]

Hi Paul,
I just posted an article on how to estimate FLOPS using event based sampling. You might want to take a look at it.

http://origin-software.intel.com/en-us/articles/estimating-flops-using-event-based-sampling-ebs/

Thanks
Levent