[Xgridagent-development] Tiger, xgridagent bugs

[Travis B. <tr...@be...>]

First of all, Tiger is finally out, so Daniel is now released from his 
NDA, and can tell us what's changed in Xgrid, and perhaps what we need 
to change in xgridagent to keep up.


I've found some bugs in xgridagent. There are also some subtle 
differences between xgridagent and Apple's GridAgent.

Bugs:

(1) xgridagent 1.0.3 hangs when sending large amounts of data (e.g. a 
couple MB tarred & gzipped) in the work directory back to the 
controller. Running with logging at level 4 (Debug with Beep messages), 
I see some XML that indicates a message of type 
XgMessageReplyTypeValue, named TaskFinalResults and seems to indicate 
that it contains the TarredWorkingDiretory. This is followed by about 
110 lines like this:
	04/29/05 15:18:46 [64784]  Sending frame of size 32700, size of window 
currently 181644
followed by a single line like this:
	04/29/05 15:18:46 [64784]  Sending frame of size 31186, size of window 
currently 181644
Then nothing--it hangs there. The Xgrid tachometer drops to 0, but the 
Status in the Xgrid client is "Job running". Files never are delivered 
to the client that submitted the job.
An identical job works fine on an OS X / Apple GridAgent agent, and a 
similar job with a much smaller output also works fine on a FreeBSD / 
xgridagent agent. I haven't figured out yet what the critical output 
size limit is yet, but I can work on that.

Any ideas on this one?

The following bugs are subtle differences between xgridagent and 
Apple's GridAgent

(2) Involves handling of literal parameters (and perhaps other 
parameters) in Custom Plug-ins. Suppose I create a custom plug-in that 
accepts 1 literal parameter, and passes it to a shell script. With OS X 
GridAgent, if the literal I pass has spaces (e.g. "-b 4"), it is passed 
to the shell script as a single parameter (it's stored in $1). With 
xgridagent, it's passed as two parameters.  There are also differences 
in how quotes within literals are handled.

(3) There seem to be other differences on whether stdout is returned 
with custom plugins, and whether files produced by a job are returned 
if there wasn't a working directory passed in as part of the custom 
plug-in. I will try to better characterize these differences.


Finally, I'd like to suggest the following behavior changes for 
xgridagent. Comments?

(4) xgridagent look somewhere other than current working directory for 
xgrid.config.xml (maybe /etc?)
(5) xgridagent should try to create "cookie" file somewhere other than 
current working directory (/tmp? /var/run?)
(6) Needs better support for multiple machines sharing the same 
xgrid.config.xml file. Either automatically include the hostname in the 
agentName, or provide an option to specify the agentName via the 
command line at launch time, to allow xgridagent to be launched by a 
script that passes in the hostname or some other unique identifier. 
This is necessary so xgridagent can be used in a cluster situation 
where all the nodes share a common drive.