Can we see how much RAM Radioss uses per CPU core during a solve?

NeilC
NeilC Altair Community Member
edited December 2020 in Community Q&A

Hi,

I need to know how much RAM each core of a CPU uses when Radioss is running HMPP and have a few questions concerning this.

There is a Memory Usage table produced at the end of a solve in the *0001.out file which indicates RAM usage, see excerpt below:

image

The table indicates memory per processor. Is this per CPU? (I run dual socket Workstations with Intel Xeon E5-2695 V3 CPU's and 256GB RAM). Is this the total written to RAM for the whole solve? Or is Radioss using 210664 MB (in the example above) when solving? so for 28 cores is this 7524 MB per core?

I would be grateful if someone could help with the understanding of how RAM is used during a solve.

The understanding affects how I need to configure hardware and local HPC or possibly utilise the Altair Cloud.

Thanks,

Neil

Tagged:

Answers

  • PaulAltair
    PaulAltair
    Altair Employee
    edited December 2020

    2 Parts to this answer, the memory is per 'process' or domain I believe (i.e. the 'np' bit of the job submission)

    However, in recent versions of RADIOSS it isn't anywhere near right, so what it represents is somewhat moot.

    The easiest way to check the actual usage is just to run 'top' on your linux machine (or process explorer or similar if under windows) to see what the job is using as it runs.

    More info:

    The memory usage reported in the 0001.out file is incorrect for SPMD/HMPP (still ok for SMP ),

    See below for a job run SPMD (-np 2) RADIOSS claims 82GB (per process) 162GB total... whereas same job run SMP (-nt 2) RADIOSS reports a rather more reasonable 77MB (the actual value)

    image

    This hasn't been correct for a few versions now, for your example Neil, it is stating you are using 181GB Average per processor which has to be wrong, your job would have crashed if that were truly the case since you only have 256GB total at best to play with.

    The memory reporting in .out did used to work ok for SPMD/HMPP in older versions, but sometime around 2018 version it got broken, I'd actually forgotten it was broken until I read your question, in practice it doesn't really matter other than an annoyance as RADIOSS is very very rarely memory limited.

    An example here comparing an older version with 2020, memory usage stats for same job , run as -NP 2 in version 11 (on left) and version 2020 on the right, the v11 numbers are right,  v2020 are very much not (again, my machine doesn't have 82Gb, let alone 162Gb)

    image

Categories