Up: IPP at the ROE

Running Simtest

As you can probably guess, Simtest is a suite of tests that, if nothing else, ensures you've got a workable IPP installed.

Requisite to running Simtest, you will need to have built IPP (IppInstallation) and initialised the database (IppInitialisingDatabase).

Configuration files

First, I will describe intallation/modification to configuration files. Probably I should have this step somewhere else, like in IPP installation.

1. cp $LOCAL/ipp/default.lin64/share/ippconfig/ipprc.config ~/.ipprc
2. Edit ~/.ipprc as you see fit. Look particularly at the PATH.
3. cp /home/ert/.ptolemyrc ~/
4. Edit ~/.ptolemyrc as you see fit. Look particularly at CONFDIR.
5. mkdir ~/ippconfig
6. cp /home/ert/ippconfig/site.config ~/ippconfig/
7. Edit ~/ippconfig/site.config appropriately. Take particular attention to DATAPATH, DBSERVER, DBNAME, DBUSER, DBPASSWORD, the last four of which you came up with when you initialised the database (IppInitialisingDatabase). Also check PSASTRO.CATDIRS. The directories don't need to exist, but they must be createable by you.

The first step, installation of a .ipprc file, seems wholly unnecessary (since the information already exists in the and is available to the code in a predictable location). Perhaps a critial review of the configuration settings would make it unnecessary. But there you go.

Through DVO, IPP has a provenance in Ohana, a collection of tools written for CFHT. Ohana tools look for user configuration in ~/.ptolemyrc. There is no master copy. It seems to just get handed down from user to user. Take mine.

site.config requires more attention. Anything under "## Stuff from Chris Waters ##" can actually be ignored for Simtest. But it might be needed in the future.

Preparing Simtest files

Simtest isn't quite in sync with changes to IPP, so it needs a bit of tweaking.

• Edit $LOCAL/ipp/default.lin64/share/ippconfig/simtest/psastro.config
• Everywhere it says *_SYNTH change to SYNTH.* (eg r_SYNTH => SYNTH.r)

Ensure you can SSH to your local client

The pantasks controller needs to be able to ssh to all nodes capable of processing data, including the local node.

Ensure it works.

• ssh localhost
• ssh hostname

SSH'ing to the local host is particularly problematic on systems with shared home directories (like the IfA) where the localhost entry in ~/.ssh/known_hosts can correspond to any machine on the network. It may need to be edited out.

Running Simtest

Simtest is run in the pantasks environment, sorta like running IDL .pro scripts in IDL.

• mkdir /wherever/you/want/to/work/simtest
• cd /wherever/you/want/to/work/simtest
• psconfig default
• pantasks
• pantasks: module simtest.pro
• pantasks: simtest testdb localhost new
• Answer YES to starting from scratch
• Supply ipp's password: aloha
• Wait for the fireworks....

pantasks will take a couple of minutes to generate the data on which to process. Then it will return to the pantasks prompt while the tasks are run in the background for the next couple of hours.

• Typing status will give the state of all the processes

Parallelisation

In IPP, Parallelisation can either mean speeding up a single code by use of threading, or running multiple tasks at the same time on the same hardware (trivial paralellisation).

Let's start with trivial parallelisation.

In pantasks, the queue controller can be accessed via the controller command. To add a machine to the list of machines to which tasks can be assigned:

• pantasks: controller host add HostToAdd

If the host is multi-core, then you can add the host many times (for the rest of the examples, HostToAdd is eday):

• pantasks: controller host add eday
• pantasks: controller host add eday
• ...

However, if the tasks are I/O intensive, then having all cores running separate tasks might not be optimal. You may just want to get the code to run in parallel with threads

To let the controller know that the host can handle multiple threads:

• pantasks: controller host add eday -threads 4

Prior to running simtest (which calls controller on its own), the number of threads to run is set by:

• pantasks: $SIMTEST_THREADS = 4

Hosts can be added before or after the jobs are submitted.

There is also a $PARALLEL variable, but I think it is just a boolean (0 or 1)

Paul Price recommends for 8-way systems to set the number of threads to something like 4 and then let the node run 3 separate processes. Results may vary.

-- EricTittley - 2009-11-06