Pyrolysis Simulation Time

[johodges]

Describe the bug
I was revisiting our internal kinetics optimization routine to integrate optimization of effective battery properties using a similar approach. I noticed the run time of the solid phase only simulations are higher in the latest versions of FDS in the cases used in the optimization. We had a similar discussion on this a few years ago with a difference between FDS 6.2 and 6.7 in #6848. The change made in 6.7.1 in the calculation of temperature dependent material properties brought the runtime down to levels in line with version 6.2.

The attached case runs ~50% slower on the latest FDS than on 6.7.1. On my local machine, the timing for 6.7.1 is ~4 seconds and for the release version of 6.9.1 is ~6 seconds. In my head I was expecting 6.9.1 to run a bit faster since we removed the overhead associated with openmp in the version I was benchmarking. In terms of solid phase only optimization, it may not be too big of a deal, but I am wondering if the difference is also having an impact in HT3D calculations where we are using a more refined grid in the solid phase.

Desktop (please complete the following information):

• OS: Windows 11
• Version: FDS 6.9.1 and 6.7.1

Additional context

coneTest2_s6.txt

[johodges]

I am looking into this a bit more. There were some syntax changes to the solid phase pyrolysis routine between versions 6.7.7 and 6.7.8. I am wondering if the changes in syntax are resulting in the difference in runtime since what is being requested by the solid phase solver is not the same. I'll let you know what I find.

[drjfloyd]

We have also made some recent changes in how small the timestep can get as a layer THICKNESS burns away to reduce numerical instablities.

[johodges]

How recent of change are you thinking? Here are the results of a timing study I ran with the input file above:

6.7.4 - 3.5 s
6.7.5 - 3.9 s
6.7.6 - 4.8 s
6.7.7 - 4.8 s
6.7.8 - 5.5 s
6.7.9 - 5.8 s
6.9.1 - 5.8 s

The jump in between 6.7.7 and 6.7.8 could be due to the syntax change, but that is only a part of the difference.

[drjfloyd]

what I was thinking of may have been since 6.9.1.

[rmcdermo]

I recompiled 6.7.1 and compared with latest source on spark. 6.7.1 took 8.236 s and latest took 10.734 s. Nothing obvious to me from the .out files as to why.

[drjfloyd]

For the latest source vs. Jonathan's observation of the release version of 6.9.1 could be changes made since 6.9.1 to timestepping when THICKNESS changes.

[rmcdermo]

I added an output quantity of SUBSTEPS both DEVC and BNDF and only see values of 1 for the latest. So, doesn't that mean we are taking DT_BC=1 with 1 substep for the solid phase?

[johodges]

When I compile the latest master on windows the timing is similar to the pre-compiled 6.9.1 timing for both at ~5.8 s. The 8.2 s on spark verse 5.8 s on my local machine is odd to me. Were you using the build openmp activated? What's the clock speed per core on spark?

[rmcdermo]

Not using openmp.

Intel(R) Xeon(R) Gold 6338 CPU @ 2.00GHz

[gforney]

here is info on one of the cores on spark001 . clock speed should be the same for other nodes/cores processor : 63 vendor_id : GenuineIntel cpu family : 6 model : 106 model name : Intel(R) Xeon(R) Gold 6338 CPU @ 2.00GHz

…

On Wed, Jul 17, 2024 at 4:09 PM johodges ***@***.***> wrote: When I compile the latest master on windows the timing is similar to the pre-compiled 6.9.1 timing for both at ~5.8 s. The 8.2 s on spark verse 5.8 s on my local machine is odd to me. Were you using the build openmp activated? What's the clock speed per core on spark? — Reply to this email directly, view it on GitHub <#13187 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AC6UCRXJIDYKGRDT4RV5JQ3ZM3FOXAVCNFSM6AAAAABLA237X6VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDEMZUGE3DQOBUHA> . You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>

-- Glenn Forney

[johodges]

My laptop is using an Intel Core i7-13800H Processor. The website says max turbo frequency of 5.2 GHz, but task manager shows FDS capping out at ~4.0 GHz.

[marcosvanella]

Hi Jon, I have a laptop with performance and efficiency cores, also Intel. I noted that using the openmp target the calculation was quite slower. Somehow it was not maxing out the clock speed.

…

________________________________ From: johodges ***@***.***> Sent: Wednesday, July 17, 2024 04:21 PM To: firemodels/fds ***@***.***> Cc: Subscribed ***@***.***> Subject: Re: [firemodels/fds] Pyrolysis Simulation Time (Issue #13187) My laptop is using an Intel Core i7-13800H Processor. The website says max turbo frequency of 5.2 GHz, but task manager shows FDS capping out at ~4.0 GHz. — Reply to this email directly, view it on GitHub<#13187 (comment)>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/ABY23VO54GR2LKLEHIDE64LZM3G5HAVCNFSM6AAAAABLA237X6VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDEMZUGIYDIMBQGU>. You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>

[mcgratta]

I'll take a look at this today. I want to look at the _cpu.csv file.

[mcgratta]

These are the results I get for the file above on spark:

coneTest2_s6_671.out  Time Stepping Wall Clock Time (s):        7.688
coneTest2_s6_674.out  Time Stepping Wall Clock Time (s):        7.650
coneTest2_s6_675.out  Time Stepping Wall Clock Time (s):        7.774
coneTest2_s6_676.out  Time Stepping Wall Clock Time (s):        9.963
coneTest2_s6_680.out  Time Stepping Wall Clock Time (s):        9.757
coneTest2_s6_new.out  Time Stepping Wall Clock Time (s):        9.823

It appears that there is a significant change between 6.7.5 and 6.7.6.