Now I have modified the script a bit with Robert's suggestion as below. Instead of starting it with postgres binary i have set it in conf file and starting the server with pg_ctl. I am waiting for the results,once the core dump is generated will share the details.
Thanks.
To verify that you do get a coredump, you might want to consider sending a kill -SEGV to a backend and make sure that it actually dumps core and you can find the core.
Ideally you'd actually set the coredumps to include shmem (see coredump_filter in http://man7.org/linux/man-pages/man5/core.5.html), but with 8GB shared_buffers that may not be practical. It'd be very useful if possible.
If this is wraparound-related, as it appears to be, you might get faster results by using a custom pgbench script for one or more workers that just runs txid_current() a whole lot. Or jump the server's xid space forward.
I've got a few other things on right now but I'll keep an eye out and hope for a core dump.
