So just to be clear, if I have the following script:

#SBATCH -N 2
#SBATCH -n 32

mpiexec.hydra -bind-to user:0,1,2,3,4,5,6,7,10,11,12,13,14,15,16,17 -n 32 ./my_program <args>

ranks 0-15 will be binded as such on the first node and ranks 16-31 will be the same for the second node? Or will all the even ranks be on one node and the odd on the other?

Thanks,

On Mon, May 11, 2015 at 9:27 PM, Kenneth Raffenetti <raffenet@mcs.anl.gov> wrote:
Ah, I see now the problem. I misread the first email. Your original line should work fine! The user bindings are listings of hw elements, not processes, so your binding will be applied identically on each node.

Ken


On 05/11/2015 05:03 PM, Justin Chang wrote:
Ken,

"-bind-to core" gives me the following topology:

process 0 binding: 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
process 1 binding: 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
process 2 binding: 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
process 3 binding: 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
process 4 binding: 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
process 5 binding: 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0
process 6 binding: 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0
process 7 binding: 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0
process 8 binding: 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0
process 9 binding: 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0
process 10 binding: 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0
process 11 binding: 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0
process 12 binding: 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0
process 13 binding: 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0
process 14 binding: 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0
process 15 binding: 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0

but I want this:

process 0 binding: 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
process 1 binding: 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
process 2 binding: 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
process 3 binding: 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
process 4 binding: 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
process 5 binding: 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0
process 6 binding: 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0
process 7 binding: 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0
process 8 binding: 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0
process 9 binding: 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0
process 10 binding: 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0
process 11 binding: 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0
process 12 binding: 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0
process 13 binding: 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0
process 14 binding: 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0
process 15 binding: 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0

The latter gives me better performance for my application, and I am
guessing it's because I have evenly distribute the processes among the
two sockets (sequentially). Which is why I resorted to what I had
originally with the custom binding.

Thanks,

On Mon, May 11, 2015 at 4:53 PM, Kenneth Raffenetti
<raffenet@mcs.anl.gov <mailto:raffenet@mcs.anl.gov>> wrote:

    Justin,

    Try using the "-bind-to core" option instead. It should do exactly
    what you are wanting. See this page with examples for more details
    https://wiki.mpich.org/mpich/index.php/Using_the_Hydra_Process_Manager#Process-core_Binding

    Ken


    On 05/11/2015 04:48 PM, Justin Chang wrote:

        Hello everyone,

        I am working with an HPC machine that has this configuring for a
        single
        compute node:

        Machine (64GB total)
            NUMANode L#0 (P#0 32GB)
              Socket L#0 + L3 L#0 (25MB)
                L2 L#0 (256KB) + L1d L#0 (32KB) + L1i L#0 (32KB) + Core
        L#0 + PU
        L#0 (P#0)
                L2 L#1 (256KB) + L1d L#1 (32KB) + L1i L#1 (32KB) + Core
        L#1 + PU
        L#1 (P#1)
                L2 L#2 (256KB) + L1d L#2 (32KB) + L1i L#2 (32KB) + Core
        L#2 + PU
        L#2 (P#2)
                L2 L#3 (256KB) + L1d L#3 (32KB) + L1i L#3 (32KB) + Core
        L#3 + PU
        L#3 (P#3)
                L2 L#4 (256KB) + L1d L#4 (32KB) + L1i L#4 (32KB) + Core
        L#4 + PU
        L#4 (P#4)
                L2 L#5 (256KB) + L1d L#5 (32KB) + L1i L#5 (32KB) + Core
        L#5 + PU
        L#5 (P#5)
                L2 L#6 (256KB) + L1d L#6 (32KB) + L1i L#6 (32KB) + Core
        L#6 + PU
        L#6 (P#6)
                L2 L#7 (256KB) + L1d L#7 (32KB) + L1i L#7 (32KB) + Core
        L#7 + PU
        L#7 (P#7)
                L2 L#8 (256KB) + L1d L#8 (32KB) + L1i L#8 (32KB) + Core
        L#8 + PU
        L#8 (P#8)
                L2 L#9 (256KB) + L1d L#9 (32KB) + L1i L#9 (32KB) + Core
        L#9 + PU
        L#9 (P#9)
              HostBridge L#0
                PCIBridge
                  PCI 1000:0087
                    Block L#0 "sda"
                PCIBridge
                  PCI 15b3:1003
                    Net L#1 "eth0"
                    Net L#2 "ib0"
                    OpenFabrics L#3 "mlx4_0"
                PCIBridge
                  PCI 8086:1521
                    Net L#4 "eth1"
                  PCI 8086:1521
                    Net L#5 "eth2"
                PCIBridge
                  PCI 102b:0533
                PCI 8086:1d02
            NUMANode L#1 (P#1 32GB) + Socket L#1 + L3 L#1 (25MB)
              L2 L#10 (256KB) + L1d L#10 (32KB) + L1i L#10 (32KB) + Core
        L#10 +
        PU L#10 (P#10)
              L2 L#11 (256KB) + L1d L#11 (32KB) + L1i L#11 (32KB) + Core
        L#11 +
        PU L#11 (P#11)
              L2 L#12 (256KB) + L1d L#12 (32KB) + L1i L#12 (32KB) + Core
        L#12 +
        PU L#12 (P#12)
              L2 L#13 (256KB) + L1d L#13 (32KB) + L1i L#13 (32KB) + Core
        L#13 +
        PU L#13 (P#13)
              L2 L#14 (256KB) + L1d L#14 (32KB) + L1i L#14 (32KB) + Core
        L#14 +
        PU L#14 (P#14)
              L2 L#15 (256KB) + L1d L#15 (32KB) + L1i L#15 (32KB) + Core
        L#15 +
        PU L#15 (P#15)
              L2 L#16 (256KB) + L1d L#16 (32KB) + L1i L#16 (32KB) + Core
        L#16 +
        PU L#16 (P#16)
              L2 L#17 (256KB) + L1d L#17 (32KB) + L1i L#17 (32KB) + Core
        L#17 +
        PU L#17 (P#17)
              L2 L#18 (256KB) + L1d L#18 (32KB) + L1i L#18 (32KB) + Core
        L#18 +
        PU L#18 (P#18)
              L2 L#19 (256KB) + L1d L#19 (32KB) + L1i L#19 (32KB) + Core
        L#19 +
        PU L#19 (P#19)

        If I ran my program with 16 processes, I would have the follow
        batch script:

        #!/bin/bash
        #SBATCH -N 1
        #SBATCH -n 20
        #SBATCH -t 0-09:00
        #SBATCH -o output.txt

        mpiexec.hydra -bind-to
        user:0,1,2,3,4,5,6,7,10,11,12,13,14,15,16,17 -n
        16 ./my_program <args>

        This would give me decent speedup. However, what if I want to use 32
        processes? Since each node only has 20 cores I would need
        #SBATCH -N 2
        and #SBATCH -n 40. However, I want ranks 0-15 and 16-31 to have
        the same
        mapping as above but on different compute nodes, so how would I
        do this?
        Or would the above line work so long as I have a multiple of 16
        processes?

        Thanks,

        --
        Justin Chang
        PhD Candidate, Civil Engineering - Computational Sciences
        University of Houston, Department of Civil and Environmental
        Engineering
        Houston, TX 77004
        (512) 963-3262 <tel:%28512%29%20963-3262>


        _______________________________________________
        discuss mailing list discuss@mpich.org <mailto:discuss@mpich.org>
        To manage subscription options or unsubscribe:
        https://lists.mpich.org/mailman/listinfo/discuss

    _______________________________________________
    discuss mailing list discuss@mpich.org <mailto:discuss@mpich.org>
    To manage subscription options or unsubscribe:
    https://lists.mpich.org/mailman/listinfo/discuss




--
Justin Chang
PhD Candidate, Civil Engineering - Computational Sciences
University of Houston, Department of Civil and Environmental Engineering
Houston, TX 77004
(512) 963-3262


_______________________________________________
discuss mailing list     discuss@mpich.org
To manage subscription options or unsubscribe:
https://lists.mpich.org/mailman/listinfo/discuss

_______________________________________________
discuss mailing list     discuss@mpich.org
To manage subscription options or unsubscribe:
https://lists.mpich.org/mailman/listinfo/discuss



--
Justin Chang
PhD Candidate, Civil Engineering - Computational Sciences
University of Houston, Department of Civil and Environmental Engineering
Houston, TX 77004
(512) 963-3262