First make sure that crontab always shows poller.php. This program will either call cmd.php, the PHP based poller _or_ cactid, the fast alternative, written in C. Define the poller you're using at Settings, Poller. Cactid has to be implemented seperately, it does not come with cacti by default.
Now, clear ./log/cacti.log (or rename it to get a fresh start)
Then, change Settings, Poller Logging Level to DEBUG for _one_ polling cycle. You may rename this log as well to avoid more stuff added to it with subsequent polling cycles.
Now, find the host/data source in question. The Host[<id>] is given numerically, the <id> being a specific number for that host. Find this <id> from the Devices menue when editing the host: The url contains a string like
id=<id>
Check, whether the output is as expected. If not, check your script (e.g. /full/path/to/perl). If ok, proceed to next step
This procedure may be replaced by running the poller manually for the failing host only. To do so, you need the <id>, again. If you're using cmd.php, set the DEBUG logging level as defined above and run
php -q cmd.php <id> <id>
If you're using cactid, you may override logging level when calling the poller:
./cactid --verbosity=5 <id> <id>
All output is printed to STDOUT in both cases. This procdure allows for repeated tests without waiting for the next polling interval. And there's no need to manually search for the failing host between hundreds of lines of output.