Fluent/CFX & Win OS: 30% slow down with MS HPC Pack 2016

SLC · January 31, 2020, 16:26

I’m posting this in the general forum as opposed to an ANSYS related one because I suspect this is an issue with the job scheduler found in the Microsoft HPC Pack 2016.

HPC Pack 2016 is based on MS MPI. That in itself is not a problem, as it is a modern MPI platform that supports fast interconnects.

I can run both CFX and Fluent locally on a compute node (dual socket, 16 core CPUs for a total of 32 cores) using Intel MPI, Platform/IBM MPI, or MS MPI with great success, giving me the expected performance. This is running Windows Server 2019.

However, as soon as I run via the job scheduler in HPC Pack 2016 on that same node, my CFX and Fluent runs see a 30% slow down. This slow down is the same if I’m running directly on the head node (with the head node also configured as a compute node), or if I’m running on a separate compute node connected via HDR infiniband.

If I, for example, run Fluent directly on the node, and start the session using msmpi, I get 100 % performance. If I start the session using the job scheduler and connect to a head node (also using msmpi) but run the simulation on that same compute node,, the simulation speed drops by 30%.

Has anyone encountered the similar problem?

SLC · February 1, 2020, 13:32

I appear to have solved my issue. I suspected it relates to affinity settings for the MS MPI, something which we can’t easily change when submitting runs to the Job Scheduler through either Fluent Launcher or CFX Solver Manager.

The quick fix is to ensure that the “exclusive” flag in the HPC Pack 2016 job template is set to “true.”. Then you get full performance as expected

January 31, 2020, 16:26	Fluent/CFX & Win OS: 30% slow down with MS HPC Pack 2016	#1
SLC Member Join Date: Jul 2011 Posts: 53 Rep Power: 15	I’m posting this in the general forum as opposed to an ANSYS related one because I suspect this is an issue with the job scheduler found in the Microsoft HPC Pack 2016. HPC Pack 2016 is based on MS MPI. That in itself is not a problem, as it is a modern MPI platform that supports fast interconnects. I can run both CFX and Fluent locally on a compute node (dual socket, 16 core CPUs for a total of 32 cores) using Intel MPI, Platform/IBM MPI, or MS MPI with great success, giving me the expected performance. This is running Windows Server 2019. However, as soon as I run via the job scheduler in HPC Pack 2016 on that same node, my CFX and Fluent runs see a 30% slow down. This slow down is the same if I’m running directly on the head node (with the head node also configured as a compute node), or if I’m running on a separate compute node connected via HDR infiniband. If I, for example, run Fluent directly on the node, and start the session using msmpi, I get 100 % performance. If I start the session using the job scheduler and connect to a head node (also using msmpi) but run the simulation on that same compute node,, the simulation speed drops by 30%. Has anyone encountered the similar problem?

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
Ansys Licence Serve on Ubuntu 16.04 LTS	david.pasquale	ANSYS	2	January 20, 2017 12:52
Slow Cases using HPC and PBS job scheduler	Alhasan	OpenFOAM Running, Solving & CFD	5	February 1, 2015 12:07
need MS HPC Pack to run in "parallel" mode	Chris Lee	SU2	0	November 24, 2014 21:05
Can we merge HPC Pack licenses?	Phillamon	FLUENT	0	January 24, 2014 03:59
Microsoft HPC Pack 2008 Tool Pack (LINPACK)	jemyungcha	Hardware	1	October 22, 2011 19:21

February 1, 2020, 13:32		#2
SLC Member Join Date: Jul 2011 Posts: 53 Rep Power: 15	I appear to have solved my issue. I suspected it relates to affinity settings for the MS MPI, something which we can’t easily change when submitting runs to the Job Scheduler through either Fluent Launcher or CFX Solver Manager. The quick fix is to ensure that the “exclusive” flag in the HPC Pack 2016 job template is set to “true.”. Then you get full performance as expected