Inderpreet Chopra wrote:
Thanks Martin for your help and support.
- According to my openion, communication through local scheduler can be
problem, if the scheduler dies before handing the jobs state to the
GRAM. In this way, the jobs will be executed, but the client will remain
thinking that the jobs are getting executed or are timed out.
Yes, but as far as i know, Gram was designed to be an interface to a local
resource manager. I think you'd have to duplicate a lot functionality and
complexity already provided by an LRM if you wanted to monitor jobs the way
you suggest.
What if your monitoring daemons on one of the cluster nodes fail, or die?
I think at some point you'll always have to deal with failures, and leveraging
the monitoring capabilities of the existing LRM's doesn't sound like a bad
compromise to me.
- Another thing you mentioned is that GRAM instance is no longer
available in grid. That i am not able to understand. Can you please
throw some more detail regarding this.
A Grid can consist of more than one site, e.g. 5 machines, located in
different networks run be different resource providers, all offering
services like Gram, RFT, GridFTP, MDS, etc.
If one of them goes down the Grid itself is not dead, there are still
the other 4 resources available. Just this particular GRAM instance
is no longer available for use by clients.
- Martin, Please correct me if i am wrong in following flow: (this can
be most generic flow for jobs submission to completion)
Java WS Core --> GRAM ---> Scheduler----> Job Execution nodes
Yes, looks ok to me.
So here WS Core can provide the interface or a way for the client to
submit the job.
GRAM, helps in submitting, monitoring and canceling the jobs.
Scheduler (PBS, Condor, etc) to schedule the jobs to different nodes.
- There is one component called MDS. Where that fits into the above flow.
ws-gram registers information about the supported local resource managers
in WS-MDS. A client can get certain information about a cluster interfaced
by ws-gram by looking up information in MDS.
For more information i suggest this link:
http://www.globus.org/toolkit/docs/4.2/4.2.1/info/#info
- I want to start implementing my ideas, to further improve the globus
processes. So, How to set the devlopment enviornment under my eclipse.
Is there some tutorial related to that. I have read, Programmers guide
for globus, but find that it gives us only information necessary to
build the webservice.
While studying, i found that there is eclipse plugin GDTE . But i am not
sure that will i be able to build new component on the existing globus
component (esp. over GRAM) using this.
What will be the correct way to start toying with the existing components.
I guess there is not something like one correct way. I personally use Eclipse,
but you can also use vi.
I do it by creating a CVS project, using ws-gram as module, adding all Java
source directories to the source in the "Java build Path" and adding a user
library that contains all GT jars.
Martin
Thanks & Regards
Inderpreet
Research Scholar
TU, Patiala
On Thu, Jan 15, 2009 at 4:48 AM, Martin Feller <[email protected]
<mailto:[email protected]>> wrote:
Inderpreet Chopra wrote:
Thanks Martin.
Please see my comments inline.
On Wed, Jan 14, 2009 at 2:49 AM, Martin Feller
<[email protected] <mailto:[email protected]>
<mailto:[email protected] <mailto:[email protected]>>> wrote:
Inderpreet Chopra wrote:
Hi all
I am working on the fault-tolerance and the security
aspects of
Grid. I want to work on these two aspects in globus .
For this I want to know about the internal structure of
Globus
and its components.
Do you mean Globus or the job management component in Globus?
I am consider about job management component only. Thanks for
providing the Execution management links, that clears many of my
doubts.
Is Globus is using some Async or sync queues for taking
request
for executing the jobs?
I'm not entirely sure what you mean by that. A client does
not hold
a connection
open until a job finished processing completely, only until a job
resource has
been created on the server-side, which is used so that a
client can
refer
to his job. Once the job resource has been created the
processing starts
on the server-side, without client-interaction.
What here i want to ask is that, client requests the GRAM for
the job execution. If there are few clients, then GRAM will
handle the requests from all the clients. But what will happen
if thousands of client try to approach the grid system for job
execution.
So i want to ask is that, is GRAM maintains some queues in which
it pushes the requests if large number of them comes at one time.
There is a queue (not in ws-gram, but in the underlying Java WS Core),
that takes incoming requests for all services. Server threads pick
requests
from the queue and the desired service is called, and finally the
response
is sent back to the client.
If the rate of incoming requests is higher than the rate of requests
being
processed, the queue fills up, and at a certain point there will be
connection timeouts. My guess is that most servers work that way and
that
the problem of 'Denial of Service' is not yet solved.
What if some node fails in between means is there some way to
recover the pending task being stuck in the erronous node?
What do you mean by "some node"? A client, the globus-server,
or a
machine of
a cluster which was picked by the local resource manager
(like PBS,
Condor) to
execute the job?
Here by node i means:
- GRAM itself fails after taking the request from the client.
Then as far as i read, the grid will be dead.
I'd rather say this GRAM instance is no longer available in the grid,
and not the grid is dead.
-The scheduling node fails, but it assigns jobs before going
into the dead state. So my question is that, Is GRAM communicate
with the execution nodes ( nodes that are actually assigned the
execution of job) directly or through the scheduler to get the
current status of the job state. If it communicate through the
scheduler, then there is problem.
GRAM gets information about the jobs via the local resource manager
(scheduler),
it does not communicate with the execution nodes directly. More
precisely, a
program (the Scheduler Event Generator (SEG)) scans the logfile of
the local
resource manager and forwards the information about jobs to GRAM
(GRAM4/ws-gram).
For GRAM2/Pre-WS-GRAM there is also periodical polling using the job
status
query commands of the local resource manager.
Why is communication via the local resource manager necessarily a
problem?
Martin
Can anyone please guide me in getting answer to my
questions and
also some documents describing globus internal architecture.
The following webpage gives an overview over key concepts
gram (job
management):
http://www.globus.org/toolkit/docs/4.2/4.2.1/execution/key/#executionKey
A starting point for more information about gram is this:
http://www.globus.org/toolkit/docs/4.2/4.2.1/execution/
Martin
Hey i am all new with Globus, so might be all these will be
stupid questions for you, but i want to clear all doubts before
actually starting working. Some leftovers that are still making
me awake are:
- If according to Job Lifetime limit, the job does not finish,
GRAM will cancel the job. Is there any means with which we can
resubmit the job automatically? I guess we need to write some
custom GRAM like component.
- What is actual way of processing single job. I guess it should
be distributed to different nodes rather than, the complete job
executed on single node. So if that is case, how GRAM manages
the response from different nodes and combine to reply to client?
Also any suggestions related to my area of intrest i.e fault
tolerance and security in globus. What is still pending
that i
can take and work upon.
Regards,
Inderpreet
Inderpreet
Research Scholar
TU, Patiala
