Hi,
The Wiki (as well as the documentation) still needs a lot of work. The
project got approved as a Globus Incubator Project just in the last
week, so the Wiki site is fresh off the press. About the installation
instructions, this link has some rough guidelines:
http://dev.globus.org/wiki/Incubator/Falkon/How_to_Run_Falkon. There is
also a README file (probably also outdated) once you donwload the source
tree.
I will make an effort to bring all the instructions up to date, and
update the Wiki. In the meantime, I'd encourage you to sign up for the
falkon-user mailing list, where we'll make general announcements (such
as major updated to the Wiki).
Also, see below for some more comments and answers...
Jan Ploski wrote:
Ioan Raicu <[EMAIL PROTECTED]> schrieb am 11/13/2007 03:53:55 AM:
Jan,
Falkon runs in user space (not system administrator space). The
pre-requisites to compile Falkon are Java 1.4 and ANT, and to run Falkon
are Java 1.4. The rest of the needed software is all bundled in the
Falkon SVN repository. The rest of the software stack is: 1) GT4 Java
WS-Core container, 2) ploticus (for graphs), 3) 500 line web server (to
view graphs remotely), and 4) the various pieces of the Falkon framework
(service code, worker code, command line client, GUI monitor). So,
basically, the end user just needs Java and ANT, which are both
downloadable and installable in a matter of 10~20 min.
Ioan,
I am not sure how clean (or up-to-date) the installation instructions
are,
"(There is currently no text in this page)"... Now that you posted the
good stuff here, please put it in the Wiki for a good start! Better yet
would be a step-by-step tutorial for a potential user, still better a
screencast (although it may be an overkill if the installation is simple).
Indeed, you may need several tutorials for different user groups.
but basically I (as the main developer of Falkon) can get the
entire software stack from SVN in 6.10 minutes (from a LAN connected
machine), compile in 1.01 min, configure things to suit the environment
in a a matter of minutes (2 configuration files control everything), and
be up and running on the order of O(10 min). If you need to get Java
and ANT as well, make it O(30 min). If you want to use security and
need to setup your security credentials (which you would have to as well
for other services such as GRAM4), then that could further add to the
time it takes to get up and running, but that is out of my control.
Do I have to deploy new WSRF services into an existing Globus container or
is Falkon just an out-of-process client?
Falkon is a web service that needs to be deployed in a GT4 container.
If you already have one, then you can use your existing container. If
not, you can use the included GT4 container.
Service deployment cannot be done
by end-users (as opposed to site administrators).
End-users can certainly deploy services in a end-user maintained
container, which is what I am proposing.
Same thing with
installing processes on a cluster head node.
You should be able to install processes on the head node for small scale
tests. For longer term testing, you probably want to either use another
compute node or you might want to use a dedicated node. For example, I
normally run the service on a dedicated machine that is on the outside
of the Grid, and I only run the provisioner on the login nodes of a
site; the provisioner is a simple Java client that is responsible for
allocating resource via GRAM4. The provisioner incurs minimal load on
the login nodes as it only periodically checks the service queues and if
needed, allocates resources via GRAM.
You should make it clear in
your installation guide whether and at which point administrator
privileges are necessary.
At no point do you need administrator privileges. Everything can be
installed as a regular user.
I did not quite understand what you mean by that
it runs in "user space" - "not a kernel thread", "not running under root
uid", "no root necessary to set up", "running under uid of a cluster
end-user" or some combination of these?
Completely user space, no root access necessary.
The biggest challenge for new applications that want to use Falkon will
be to either wrap the existing command line client, or to change the
command line client to suit their needs. However, if an application
already has support for GRAM, the transition to Falkon should be
relatively straight forward as the semantics are similar (define job
description, submit, wait for notification of completion).
I hope its clear that the middleware at a grid site does not need to
change at the site level.
It is certainly not apparent from your web site. In fact, the impression I
got from reading the SC07 presentation was that it's a real boon for
Swift/Karajan users, which invokes thoughts like "ok, so where do I
start... by reading Swift/Karajan documentation, what fun...".
Fortunately, this no longer seems to be true.
Swift is a vehicle for Falkon to gain access to a good set of
applications that do not have to be modified explicitly to use Falkon
(as Swift was already modified to submit jobs to Falkon). Falkon is
actually being used in a separate astronomy application that was built
from the ground up based on Falkon (no Swift involved). The key to see
how you can use Falkon is the command line client that is in the SVN
archive.
Falkon will be a user level process that runs
on behalf of the user who configured and started it, will allocate
resources via GRAM4, dispatch tasks via WS calls directly to compute
resources (out of band of GRAM), etc. Once its configured at a site
with the GRAM4 server location, number of resources needed (upper and
lower bounds), the time resources are needed for, etc, starting and
stopping Falkon is a 1 command invocation. Once started, Falkon will
monitor the Falkon queue, and increase the number of resources when the
queue grows, and shrink the number of resources when the queue is empty
(all within the bounds specified in the configuration of Falkon).
Please post this information on your site, too.
Right, I'll be updating the Wiki this week (hopefully).
If you happen to try Falkon out, I'd be happy to hear back feedback
about how easy it was to setup to run some simple sleep jobs, and even
more how easy/difficult it was to use the client API to integrate into
your existing application!
You have certainly piqued my interest. I will get back to you when I try
it (December).
Great! Looking forward to hearing back on your experience.
Ioan
Regards,
Jan Ploski
