Anindya Bose wrote:
Your Plan Viz shows 3 dominant operators with 45, 32 and 0.62 mili seconds and then overall run time is 333 seconds ( slightly more than 5 mins ) .. I am confused here. Where the time being spent then ?
That's exactly the question that needs to be answered here.
It's not about the execution plan, the data distribution or the way SAP HANA computes joins.
By looking at the PlanViz overview statistics it's already clear, that the instrumented runtime accounting does not cover the long total response time.
If the system is not massively burning CPU during the 5 mins of wall clock time, then it must be waiting for something.
So, checking the state of the threads involved in the query execution (also looking at the stack traces) can be a good first step. If the threads are hanging in some os calls or waiting for low level locks (mutex, futex, semaphores is what you'd typically see in the SAP HANA stack traces), then the next step is to find who's holding those locks.
Another angle for approaching this problem can be to check the OS/HW setup. This a XEN virtual system, so things like CPU assignment/affinity, hyperthreading and transparent huge page setup needs to be checked on all involved levels of the system setup.