questions about LocalMapdlPool and launch_mapdl

[shixinghari]

I think running pool can save the time comsumption, but things are totally different as below, I wonder if I done something wrong

and unlike the example in (https://courses.ansys.com/index.php/courses/pymapdl-to-create-physics-based-datasets-for-ml-ai-applications/lessons/creating-datasets-for-simulation-models-with-pymapdl-lesson-4/)
in my fem model, my CPU takes just 20% when use

pool = LocalMapdlPool(4, nproc=20, additional_switches='-dmp')
thetas = rnd_k(50)
out = pool.map(fem1, thetas, progress_bar=True, wait=True)

just as same when I run
mapdl = launch_mapdl()
But when I run the example in lesson-4, the rate of CPU is much higher.

The situation makes me confused, my study is about bayesian model updating, need to do modal analysis many times, so I tryed LocalMapdlPool to speed up my study, but things is totally different, looking forward to help

about my fem model, I show below

def fem1(mapdl, parameters):
    """
    这是平面四层两跨实验室钢结构有限元模型函数\n
    参考[1]基于贝叶斯方法的有限元模型修正研究_张建新.pdf[J].\n
    输入四个参数作为整体，输出频率\n
    :param mapdl: 调用ansys进行并行化必须的参数
    :param parameters: 该模型需要的参数，分别是柱的刚度和梁的质量
    :return: freq-四阶自振频率
    """

    if len(parameters) == 1:
        parameters = parameters[0]
    error = 'the dimension of parameter is wrong, this fem requires 4, but got {}'.format(len(parameters))
    assert len(parameters) == 4, error

    mapdl.clear()
    # 前处理，模型属性
    mapdl.prep7()
    x = [1] * 20
    #  调整弹性模量
    x[0] = parameters[0]  # 一层柱
    x[5] = parameters[0]  # 二层柱
    x[10] = parameters[1]  # 三层柱
    x[15] = parameters[1]  # 四层柱
    y = [1] * 20
    #  调整梁的密度
    y[3] = parameters[2]  # 一层梁
    y[8] = parameters[2]  # 二层梁
    y[13] = parameters[3]  # 三层梁
    y[18] = parameters[3]  # 四层梁
    for i in range(20):
        mapdl.et(i + 1, 'beam189')
        mapdl.mp("EX", i + 1, x[i] * 2.1e11)  # N/mm2
        mapdl.mp('DENS', i + 1, y[i] * 7.8e3)  # kg/mm3
        mapdl.mp("PRXY", i + 1, 0.3)  # Poisson's ratio
        mapdl.sectype(i + 1, "BEAM", "RECT")
        mapdl.secdata(0.065, 4e-3)

    # 创建节点
    for i in range(5):
        mapdl.k(3 * i + 1, -0.350, 0.350 * i, 0)
        mapdl.k(3 * i + 2, 0, 0.350 * i, 0)
        mapdl.k(3 * i + 3, 0.350, 0.350 * i, 0)
    mapdl.k(16, 0.700, 0.700, 0)
    mapdl.k(17, 0, 0.7700, 0)

    # 创建线
    for i in range(4):
        mapdl.l(3 * i + 1, 3 * i + 4)
        mapdl.l(3 * i + 2, 3 * i + 5)
        mapdl.l(3 * i + 3, 3 * i + 6)
        mapdl.l(3 * i + 4, 3 * i + 5)
        mapdl.l(3 * i + 5, 3 * i + 6)

    # 将材料、实常数、单元类型、、方向点、截面类型赋予此前选中的线模型
    for i in range(4):
        # 选中各层柱线,1一层柱，6二层柱，11三层柱，16四层柱
        mapdl.lsel("S", "LINE", "", [5 * i + 1, 5 * i + 2, 5 * i + 3])
        #  将材料、实常数、单元类型、、方向点、截面类型赋予此前选中的线模型
        mapdl.latt(5 * i + 1, 5 * i + 1, 5 * i + 1, "", 16, 5 * i + 1)
        #  选中各层梁线
        mapdl.lsel("S", "LINE", "", [5 * i + 4, 5 * i + 5])
        mapdl.latt(5 * i + 4, 5 * i + 4, 5 * i + 4, "", 17, 5 * i + 4)

    # 网格划分
    mapdl.allsel()
    mapdl.lesize("all", "", "", 1)
    mapdl.lmesh("all")

    # 执行分析
    mapdl.slashsolu()
    mapdl.ksel('all')
    mapdl.dk('all', 'uz')
    mapdl.dk(1, "all")
    mapdl.dk(2, "all")
    mapdl.dk(3, "all")
    mapdl.modal_analysis(nmode=4, freqb=1)
    # ==============获取结果
    """disps = []
    for i in range(3):
        displacement = mapdl.result.nodal_displacement(i)
        disp = displacement[1][:, 0]
        disp = disp / np.linalg.norm(disp, ord=2)
        disps.append(disp)"""
    mapdl.post1()
    freq = np.zeros(4)
    for i in range(4):
        mapdl.set(1, i + 1)
        freq[i] = mapdl.post_processing.freq
    mapdl.finish()
    return freq

[shixinghari]

By the way, I just run LocalMapdlPool and launch_mapdl in the example showed in https://courses.ansys.com/index.php/courses/pymapdl-to-create-physics-based-datasets-for-ml-ai-applications/lessons/creating-datasets-for-simulation-models-with-pymapdl-lesson-4/
I cut the inputs size from 3000+ to 100, the result shows that launch_mapdl is faster, takes 54s, while LocalMapdlPool takes 120s

First, the question is what is the advantage of using LocalMapdlPool while it's slower than launch_mapdl
Second, about bayesian model updating, I need to run my fem model thousands of time. Is there any methods to reduce the time?

[germa89]

Hi @shixinghari

Quick thought without diving to much yet. Can you try same configuration regarding number of processors?
It does not seem a very big model so you can use nproc=1 for both MapdlPool and launch_mapdl.
Same with the -dmp flag, I would remove it for a fair comparison.

Then post the new times here, please

[shixinghari]

Thank you for your reply! @germa89

Due to the time zone difference, I just woke up. I apologize for the delayed response.

Just tryed as you said, things never changed, but it semms that running MapdlPool with 1 core is faster than MapdlPool with 20 cores.

When I tried to use the method of controlling variables pool = LocalMapdlPool(4, nproc=20), I encountered a small error.

I tryed again, without error, time = 284.5947530269623

t1=time.time()
pool = LocalMapdlPool(4, nproc=20)
thetas = rnd_k(50)
out = pool.map(fem1, thetas, progress_bar=True, wait=True)
t2=time.time()
print(t2-t1)```

[germa89]

Sorry, this slide on my inbox.

Yes, running with 20 cores (nproc=20) is slower to start than using only 1 (nproc=1). Furthermore, you might hit license limitations if you are trying to use more than you have.

If you are using the student version, you can only have one MAPDL instance (up to 4 cores I think).

[germa89]

Please, do double check you can start four MAPDL instances with 20 cores each (licensing wise, it might be a lot)

[shixinghari]

@germa89
Hello, I wonder if it's due to my own situation that makes the localmapdlpool slower than launch_mapdl.
So, I wonder if localmapdlpool runs faster than launch_mapdl in your device?

[germa89]

MapdlPool uses launch_mapdl in the background to start the MAPDL instances, so it will be always "slightly" (almost neglectable) slower than using launch_mapdl directly.

[mikerife]

@shixinghari
Could you please use MapdlPool to start one instance of MAPDL aaS with 20 cores (DMP parallel) and have the Windows Task Manager open showing the Details tab. When the 20 instances of ansys.exe start up do they seem to do so all at once? Or one at a time with some pause between each one starting? If there is a pause then this could be a security/scanning software interferring with the start up procedure.
Mike

[shixinghari]

Thanks for your reply!
I was not quite understanding about how to do as you said.
So I will repeat what I have tried.

1, when I use pool = LocalMapdlPool(4, nproc=4, additional_switches = '-dmp') , the ANSYS instance started one by one, this 4 instance takes me 28s, while I use pool = LocalMapdlPool(4, nproc=4, additional_switches = '-smp'), the ANSYS instance started one by one, this 4 instance takes me 4s. BUT both of them run slower in the later analyse than launch_mapdl()
2, I doubt it is because of liscence limitation, BUT when I run two launch_mapdl() in two ipynb files, it works good as usual, much faster than pool = LocalMapdlPool(2, nproc=20, additional_switches = '-dmp').
3, Then I think doubt (this could be a security/scanning software interferring), I shut down 'wdsc-virus-threat-protection', didn't work.

Since, running two launch_mapdl() in two different ipynb files, works just good and fast as usual, and germa89 said (MapdlPool uses launch_mapdl in the background to start the MAPDL instances, so it will be always "slightly" (almost neglectable) slower than using launch_mapdl directly.)

I think the problem maybe in data transmission, when ANSYS complete one task, it may sleep for a while for some reason (maybe dealing and passing the outcome) and this process takes longer in LocalMapdlPoolthan launch_mapdl().
I shut down 'wdsc-virus-threat-protection', trying to keep ANSYS from security check, but didn't work.

And I need to apologize for using some special version of ANSYS 2021R1, maybe this could be one of the reasons.

[germa89]

@shixinghari can you try to replicate this 8-cores example (#3115 (reply in thread)) and give us the times?

It should be very similar. If not, something might be wrong with your setup.

[shixinghari]

Thanks, I have try what mikelife said.
First about the time

after I repeat launching LocalMapdlPool(I can't import MapdlPool), the time of -dmp slow down, as you can see below

[shixinghari]

Second, about the Windows Task Manager,
pool = LocalMapdlPool(n_instances=4, nproc=6, additional_switches='-dmp') will launch 4*6=24 ANSYS.exe at the same time with 4 ANSYS211.exe
while pool = LocalMapdlPool(n_instances=4, nproc=6, additional_switches='-smp') will launch 4 ANSYS.exe and 4 ANSYS211.exe at the same time, seems -smp didn't care about nproc, the number of ANSYS.exe only influenced by n_instances

Third, both -smp and -dmp in LocalMapdlPool take longer time than launch_mapdl in the same analysis.
All of these 3 methods, only n_instances numbers of the CPU time of ANSYS.exe become bigger with time, as shower in the picture above, only 4 ANSYS.exe have over 1min.

[germa89]

Thanks, I have try what mikelife said. First about the time after I repeat launching LocalMapdlPool(I can't import MapdlPool), the time of -dmp slow down, as you can see below

What changed in the second screenshot with respect the first? Did you just repeat the same code?

[shixinghari]

just repeat the same code for several times
and sometimes, after the first launch and exit, pool = LocalMapdlPool(n_instances=4, nproc=6, additional_switches='-dmp') will fail to start ansys

[mikerife]

@shixinghari @germa89
Can you update to a newer version of MAPDL? At 2021 the licensing changed in a big way; the interconnect went away. So this was the first release where the applications we doing more of the client side licensing functions aside from just requesting a license. Looks like it is a known issue that the license check-out was...unstable compared to prior releases. Which then got addressed in 2022 but 2023 is better. So I don't think that there is a PyMAPDL side 'fix' for this.
Mike