Differences

This shows you the differences between two versions of the page.

Link to this comparison view

Both sides previous revision Previous revision
Next revisionBoth sides next revision
doku:gromacs [2023/05/11 15:45] – msiegeldoku:gromacs [2023/05/17 12:29] – msiegel
Line 3: Line 3:
 Our recommendation: Follow these three steps to get the fastest program: Our recommendation: Follow these three steps to get the fastest program:
  
-  - Use the most recent version of GROMACS that we provide or build your own. +  - Use the most **recent version** of GROMACS that we provide or build your own. 
-  - Use the newest Hardware: the partitions ''zen2_0256_a40x2'' or ''zen3_0512_a100x2'' on VSC-5 have plenty nodes available with 2 GPUs each. +  - Use the newest Hardware: use **1 GPU** on the partitions ''zen2_0256_a40x2'' or ''zen3_0512_a100x2'' on VSC-5. 
-  - Do some performance analysis to decide if a single GPU Node (likely) or multiple CPU Nodes via MPI (unlikely) better suits your problem.+  - Do some **performance analysis** to decide if a single GPU Node (likely) or multiple CPU Nodes via MPI (unlikely) better suits your problem.
  
-In most cases it does not make sense to run on multiple GPU nodes with MPI; weather using only one or both GPUs per node.+In most cases it does not make sense to run on multiple GPU nodes with MPI; Whether using one or two GPUs per node.
  
 ===== GPU Partition ===== ===== GPU Partition =====
Line 14: Line 14:
 this a ''partition'', described in detail at [[doku:slurm | this a ''partition'', described in detail at [[doku:slurm |
 SLURM]]. On any login node, type ''sinfo'' to get a list of the SLURM]]. On any login node, type ''sinfo'' to get a list of the
-available partitions. The partition has to be set in the batch script,+available partitions, or take a look at [[doku:vsc4_queue]] and [[doku:vsc5_queue]]. We recommend the partitions ''zen2_0256_a40x2'' or ''zen3_0512_a100x2'' on VSC-5; they have plenty nodes available with 2 GPUs each. The partition has to be set in the batch script,
 see the example below. Be aware that each partition has different see the example below. Be aware that each partition has different
 hardware, so choose the parameters accordingly. GROMACS decides mostly on its own how it wants to hardware, so choose the parameters accordingly. GROMACS decides mostly on its own how it wants to
Line 25: Line 25:
    
   * some SLURM parameters: the ''#BATCH ...'' part)   * some SLURM parameters: the ''#BATCH ...'' part)
-  * exporting environment variables: e.g. ''export CUDA_VISIBLE_DEVICES=0'' +  * export environment variables: e.g. ''export CUDA_VISIBLE_DEVICES=0'' 
-  * cleaning the environment: ''module purge'' +  * clean modules: ''module purge'' 
-  * loading modules: ''load gcc/7.3 ...'' +  * load modules: ''module load gromacs/2022.2-gcc-9.5.0-...'' 
-  * last but not least starting the program in question: ''gmx_mpi ...''+  * starting the program in question: ''gmx_mpi ...''
  
 <code bash mybatchscript.sh> <code bash mybatchscript.sh>
Line 41: Line 41:
  
 module purge module purge
-module load cuda/11.5.0-gcc-11.2.0-ao7cp7w openmpi/4.1.4-gcc-11.2.0-ub765vm python/3.8.12-gcc-11.2.0-rvq5hov gromacs/2022.2-gcc-11.2.0-4x2vwol +module load gromacs/2022.2-gcc-9.5.0-...
 gmx_mpi mdrun -s topol.tpr gmx_mpi mdrun -s topol.tpr
 </code> </code>
Line 51: Line 50:
  
  
-===== CPU / GPU Load =====+===== Performance ===== 
 + 
 + 
 +==== CPU / GPU Load ====
  
 There is a whole page dedicated to [[doku:monitoring]] the CPU and There is a whole page dedicated to [[doku:monitoring]] the CPU and
Line 58: Line 60:
  
  
-===== Performance =====+==== Short Example ====
  
 As a short example we ran ''gmx_mpi mdrun -s topol.tpr'' with As a short example we ran ''gmx_mpi mdrun -s topol.tpr'' with
 different options, where ''topol.tpl'' is just some sample topology, different options, where ''topol.tpl'' is just some sample topology,
-we don't actually care about the result. Without any options GROMACS+we don't actually care about the result, we just want to know how many **ns/day** we can get, Gromacs tells you that at the end of every run. Such a short test can be done in no time. 
 + 
 +The following table lists our 5 tests: Without any options GROMACS
 already runs fine (a). Setting the number of tasks (b) is not needed; already runs fine (a). Setting the number of tasks (b) is not needed;
 if set wrong can even slow the calculation down significantly ( c ) due if set wrong can even slow the calculation down significantly ( c ) due
-to over provisioning! We would advice to enforce pinning, in our+to over provisioning! We would advise to enforce pinning, in our
 example it does not show any effects though (d), we assume that the example it does not show any effects though (d), we assume that the
 tasks are pinned automatically already. The only further improvement tasks are pinned automatically already. The only further improvement
Line 72: Line 76:
  
 ^ # ^ cmd         ^ ns / day ^ cpu load / % ^ gpu load / % ^ notes                               ^ ^ # ^ cmd         ^ ns / day ^ cpu load / % ^ gpu load / % ^ notes                               ^
-| a | --          | 160    | 100   | 80        |                                    | +| a | ''--''          | 160    | 100   | 80        |                                    | 
-| b | -ntomp 8    | 160    | 100   | 80        |                                    | +| b | ''-ntomp 8''    | 160    | 100   | 80        |                                    | 
-| c | -ntomp 16   | 140    | 40    | 70        | gromacs warning: over provisioning | +| c | ''-ntomp 16''   | 140    | 40    | 70        | gromacs warning: over provisioning | 
-| d | -pin on     | 160    | 100   | 80        |                                    | +| d | ''-pin on''     | 160    | 100   | 80        |                                    | 
-| e | -update gpu | 170    | 100   | 90        |                                    |+| e | ''-update gpu'' | 170    | 100   | 90        |                                    | 
 + 
 + 
 +==== 7 Test Cases ==== 
 + 
 +Since Gromacs is used in many and very different ways, it makes sense to 
 +benchmark various scenarios: 
 + 
 +  - a VSC users test case (??? atoms) 
 +  - R-143a in hexane (20,248 atoms) with very high output rate 
 +  - a short RNA piece with explicit water (31,889 atoms) 
 +  - a protein inside a membrane surrounded by explicit water (80,289 atoms) 
 +  - a protein in explicit water (170,320 atoms) 
 +  - a protein membrane channel with explicit water (615,924 atoms) 
 +  - a huge virus protein (1,066,628 atoms) 
 + 
 +Take a look at the test results resembling a similar case than your application. 
 + 
 +In this chart we tested our various hardware on the 7 test cases, some recent GPUs like ''A100'' GPU Nodes on VSC-5, some belong to VSC-3 (like ''GTX1080'', ''K20''), and are not accessible anymore: 
 + 
 +<achart> 
 +{ 
 +    series: [{ 
 +        name: 'Test 1', 
 +        data: [191, 144, 128, 125, 145, 127, 92, 62, 57, 60, 57, 29, 28, 27, 17, 7.4, 7.4] 
 +    }, { 
 +        name: 'Test 2', 
 +        data: [525, 442, 449, 455, 471, 317, 228, 228, 207, 193, 152, 73, 74, 61, 46, 18, 18] 
 +    }, { 
 +        name: 'Test 3', 
 +        data: [205, 143, 164, 130, 113, 164, 103, 66, 68, 58, 48, 24, 25, 23, 14, 6.2, 6] 
 +    }, { 
 +        name: 'Test 4', 
 +        data: [463, 333, 273, 246, 229, 276, 103, 165, 170, 158, 143, 69, 67, 54, 40, 16, 16] 
 +    }, { 
 +        name: 'Test 5', 
 +        data: [168, 139, 162, 147, 131, 174, 94, 61, 59, 58, 43, 18, 18, 22, 10, 5.2, 5] 
 +    }, { 
 +        name: 'Test 6', 
 +        data: [9.6, 8.1, 16, 8.4, 9.9, 7.3, 12, 4.3, 3.1, 3.1, 4.6, 1.7, 1.7, 1.6, 1, 0.4, 0.4] 
 +    }, { 
 +        name: 'Test 7', 
 +        data: [27.2, 13, 25, 21.8, 1.4, 24.6, 18, 8.6, 8, 7.6, 8, 3.1, 3.1, 3, 1.7, 0.7, 0.7] 
 +    }], 
 +    chart: { 
 +        type: 'bar', 
 +        height: 350, 
 +        stacked: true, 
 +    }, 
 +    plotOptions: { 
 +        bar: { 
 +            horizontal: true, 
 +        }, 
 +    }, 
 +    title: { 
 +        text: 'Gromacs Benchmarks, 7 tests on various hardware' 
 +    }, 
 +    xaxis: { 
 +        categories: [ 
 +            "1x A40", 
 +            "1x RTX2080TI", 
 +            "1x A100", 
 +            "4x GTX1080 M", 
 +            "2x A40", 
 +            "8x GTX1080 M", 
 +            "2x A100", 
 +            "2x GTX1080 M", 
 +            "1x GTX1080 M", 
 +            "1x GTX1080 S", 
 +            "0x A100", 
 +            "0x GTX1080 M", 
 +            "0x A40", 
 +            "1x K20M", 
 +            "0x K20M", 
 +            "0x GTX1080  S", 
 +            "0x RTX2080TI", 
 +        ], 
 +        title: { 
 +            text: "ns / day" 
 +        }, 
 +    }, 
 +    legend: { 
 +        position: 'right', 
 +        horizontalAlign: 'left', 
 +        title: { 
 +            text: "Test #" 
 +        }, 
 +    } 
 +} 
 +</achart> 
 + 
 + 
 +==== Many GPUs ==== 
 + 
 +In most cases 1 GPU is **better** than 2 GPUs! 
 + 
 +In some cases, for example a large molecule like Test 7, you might want to run Gromacs on both GPUs. We strongly encourage you to test if you actually benefit from running with GPUs on many nodes. 
 + 
 +To find out if more GPUs mean more work done we need some math: the parallel efficiency **η** is the ratio of the [[https://en.wikipedia.org/wiki/Speedup | speedup]] factor **S(N)** and the number of processors **N**: 
 + 
 +η = S(N) / N 
 + 
 +In this chart we compare Gromacs parallel efficiency **η** of the 7 test cases with two GPU versus one GPU on VSC-5 ''A40'' and ''A100'' Nodes. Ideally two GPUs would bring a speedup **S** of 2, but also a **N** of 2, so the parallel efficiency would be 1. In reality, the speedup will be less than 1; and so will be the efficiency. An efficiency 0.45 for example means that with two GPUs, more than half the used resources were wasted! Because of the high demand on our GPU Nodes we kindly ask to test your case, and only use more than 1 GPU, if your efficiency is above 0.5. 
 + 
 +Set the number of GPUs on the node visible to Gromacs with ''export CUDA_VISIBLE_DEVICES=0'' for one GPU or ''export CUDA_VISIBLE_DEVICES=0,1'' for both GPUs. 
 + 
 +<achart> 
 +{ 
 +    series: [{ 
 +        name: '2x A40', 
 +        data: [0.38, 0.45, 0.28, 0.25, 0.39, 0.52, 0.03] 
 +    }, {    
 +        name: '2x A100', 
 +        data: [0.36, 0.25, 0.31, 0.19, 0.29, 0.38, 0.36] 
 +    }], 
 +    chart: { 
 +        type: 'bar', 
 +        height: 350, 
 +        // stacked: true, 
 +        // stackType: '100%' 
 +    }, 
 +    // plotOptions: { 
 +    //     bar: { 
 +    //         horizontal: true, 
 +    //     }, 
 +    // }, 
 +    title: { 
 +        text: 'Strong Scaling Efficiency of dual GPU' 
 +    }, 
 +    xaxis: { 
 +        categories: [ 
 +            "Test 1", 
 +            "Test 2", 
 +            "Test 3", 
 +            "Test 4", 
 +            "Test 5", 
 +            "Test 6", 
 +            "Test 7", 
 +        ], 
 +        // title: { 
 +        //     text: "Test #" 
 +        // }, 
 +    }, 
 +    yaxis: { 
 +        title: { 
 +            text: "Strong Scaling Efficiency S(N) / N" 
 +        }, 
 +    }, 
 +    legend: { 
 +        position: 'right', 
 +        horizontalAlign: 'left', 
 +    } 
 +} 
 +</achart> 
 + 
 + 
 +==== Many nodes with many GPUs ==== 
 + 
 +In most cases one node is **better** than more nodes. 
 + 
 +In some cases, for example a large molecule like Test 7, you might want to run Gromacs on 
 +multiple nodes in parallel using MPI, with multiple 
 +GPUs (one each node). We strongly encourage you to test if you 
 +actually benefit from running with GPUs on many nodes. Gromacs can perform worse on 
 +many nodes in parallel than on a single one, even considerably! 
 + 
 +Run Gromacs on multiple nodes with: 
 + 
 +<code bash> 
 +#SBATCH --nodes 2 
 +gmx mdrun ... 
 +</code> 
 + 
 +Take a look at the chapter [[doku:gromacs#Dual GPU Scaling Efficiency]] to find out if more nodes are a good idea. In the below examples the efficiency is very low, even for the large test cases 6 and 7. 
 + 
 +<achart> 
 +{ 
 +    series: [{ 
 +        name: 'Test 1', 
 +        data: [ 42.374, 40.176, 39.439, 38.252, 35.744, 30.811 ] 
 +    }, { 
 +        name: 'Test 2', 
 +        data: [ 82.513, 81.25, 84.805, 81.894, 72.589, 62.855 ] 
 +    }, { 
 +        name: 'Test 3', 
 +        data: [ 0, 0, 0, 0, 0, 0 ] 
 +    }, { 
 +        name: 'Test 4', 
 +        data: [ 0, 0, 0, 0, 0, 0 ] 
 +    }, { 
 +        name: 'Test 5', 
 +        data: [ 67.147, 76.027, 80.627, 80.903, 83.031, 68.702 ] 
 +    }, { 
 +        name: 'Test 6', 
 +        data: [ 10.612, 11.963, 10.996, 14.37, 35.482, 34.988 ] 
 +    }, { 
 +        name: 'Test 7', 
 +        data: [ 17.92, 21.604, 30.482, 37.497, 35.448, 43.254 ] 
 +    }], 
 +    chart: { 
 +        type: 'bar', 
 +        height: 350, 
 +        stacked: true, 
 +    }, 
 +    plotOptions: { 
 +        bar: { 
 +            horizontal: true, 
 +        }, 
 +    }, 
 +    title: { 
 +        text: 'Gromacs Benchmarks: Performance on multiple Nodes via MPI' 
 +    }, 
 +    xaxis: { 
 +        categories: [ 
 +            "1 Node", 
 +            "2 Nodes", 
 +            "4 Nodes", 
 +            "8 Nodes", 
 +            "16 Nodes", 
 +            "32 Nodes", 
 +        ], 
 +        title: { 
 +            text: "ns / day" 
 +        }, 
 +    }, 
 +    legend: { 
 +        position: 'right', 
 +        horizontalAlign: 'left', 
 +        title: { 
 +            text: "Test #" 
 +        }, 
 +    } 
 +} 
 +</achart> 
 + 
 + 
 +==== Many ranks on many nodes with many GPUs==== 
 + 
 +Quick summary: 
 +  * Most problems (Small): 1 or 2 ranks per node 
 +  * Large problem: 8 ranks per node 
 + 
 +If you want to run Gromacs on multiple nodes and multiple GPUs in parallel using MPI, best 
 +tell MPI how many processes should be launched on each node 
 +''--map-by ppr:1:node''. MPI would normally use many ranks, which is good for CPUs, but bad with GPUs! We encourage you to test a few number of ranks 
 +yourself with your specific application. Based on our tests listed in the following chart we 
 +recommend 1 ranks per node for most (small) problems, and only for large problems up to 8 ranks per node: 
 + 
 +<code bash> 
 +#SBATCH --nodes 2 
 +mpirun -np 16 \ 
 +       --map-by ppr:8:node \ 
 +       ... 
 +       gmx_mpi mdrun ... 
 +</code> 
 + 
 +The reason for this is that the graphics cards does more work than the CPU. Gromacs needs to copy data between 
 +different ranks on the CPUs and all GPUs, which takes more time with more ranks. Gromacs notices that and shows 
 +''Wait GPU state copy'' at the end of the log. As an example our test 
 +1 with 16 ranks on 1 node: the ''Wait GPU state copy'' amounts to 44.5% 
 +of the time spent!
  
 +<achart>
 +{
 +    series: [{
 +        name: 'Test 1',
 +        data: [ 43.644, 46.385, 32.454, 37.333, 19.084, 16.136, 4.824 ]
 +    }, {
 +        name: 'Test 2',
 +        data: [ 390.057, 138.831, 89.078, 78.769, 39.94, 35.99, 9.545 ]
 +    }, {
 +        name: 'Test 3',
 +        data: [ 82.997, 39.682, 33.176, 80.643, 48.766, 29.216, 13.972 ]
 +    }, {
 +        name: 'Test 4',
 +        data: [ 144.859, 52.099, 35.469, 96.125, 55.373, 32.502, 14.864 ]
 +    }, {
 +        name: 'Test 5',
 +        data: [ 30.174, 35.561, 39.051, 68.824, 39.012, 34.442, 10.475 ]
 +    }, {
 +        name: 'Test 6',
 +        data: [ 18.282, 10.061, 15.62, 20.889, 17.528, 16.452, 7.534 ]
 +    }, {
 +        name: 'Test 7',
 +        data: [ 26.499, 14.855, 22.433, 26.672, 21.686, 19.323, 7.879 ]
 +    }],
 +    chart: {
 +        type: 'bar',
 +        height: 350,
 +        stacked: true,
 +    },
 +    plotOptions: {
 +        bar: {
 +            horizontal: true,
 +        },
 +    },
 +    title: {
 +        text: 'Gromacs Benchmarks: Performance with various Ranks on 2 Nodes with 1 GPU each'
 +    },
 +    xaxis: {
 +        categories: [
 +            "1 Rank",
 +            "2 Ranks",
 +            "4 Ranks",
 +            "8 Ranks",
 +            "16 Ranks",
 +            "28* Ranks",
 +            "64 Ranks",
 +        ],
 +        title: {
 +            text: "ns / day"
 +        },
 +    },
 +    legend: {
 +        position: 'right',
 +        horizontalAlign: 'left',
 +        title: {
 +            text: "Test #"
 +        },
 +    }
 +}
 +</achart>
  
  
  • doku/gromacs.txt
  • Last modified: 2023/11/23 12:27
  • by msiegel