Software Archive
Read-only legacy content
17061 Discussions

Failed to explicit offload on Xeon Phi KNC

Rolly_N_
Beginner
‎03-10-2017 09:53 AM
1,327 Views

Dear all,

I am referring to the guide on Explicit offload QE to Xeon Phi KNC (7120P) here,

https://software.intel.com/en-us/articles/explicit-offload-for-quantum-espresso

I tried to follow the above steps but I failed to run pw.x (QE v5.3.0) on 2 Xeon Phi 7120P using the mpirun.sh script. The error reads:

 allocating buffers        2048        2048        1024
 on device            0
 threshold    20000000000.0000     
 allocating buffers        2048        2048        1024
 on device            0
 threshold    20000000000.0000     
offload error: cannot create buffer on device 0 (error code 14)
offload error: cannot create buffer on device 0 (error code 14)

This is how I run the script,

[qeuser@node09 ~]$ ~/mpirun/mpirun.sh -p 1 -w ~/libxphi/xphilibwrapper.sh -x ~/QE530-KNC-OL/espresso-5.3.0/bin/pw.x -i ~/rolly/AUSURF112/ausurf.in

I have already scp all the lib and bin files to each Xeon Phi 7120P and I have also compiled the libxphi lib. This is how the libxphi directory reads,

[qeuser@node09 libxphi]$ ls
build-library.sh  libmkl_proxy.so  LICENSE      README.md    xphilibmod.mod     xphilib.o          xphilib_proxy.o
clean.sh          libxphi.so       mkl_proxy.c  xphilib.f90  xphilibmod.modmic  xphilib_proxy.f90  xphilibwrapper.sh

I suppose this is okay.

However, I found it interesting that I can run a single instance on mic0 but it is very slow. This is how I did it,

[qeuser@node09 ~]$ export LD_LIBRARY_PATH=/home/qeuser/libxphi/:$LD_LIBRARY_PATH

[qeuser@node09 ~]$ LD_PRELOAD="/home/qeuser/libxphi/libxphi.so" /home/qeuser/QE530-KNC-OL/espresso-5.3.0/bin/pw.x  < /home/qeuser/rolly/AUSURF112/ausurf.in

Error messages were also produced,

 allocating buffers        2048        2048        1024
 on device            0
 threshold    20000000000.0000     
ERROR: ld.so: object '/home/qeuser/libxphi/libxphi.so' from LD_PRELOAD cannot be preloaded (cannot open shared object file): ignored.
ERROR: ld.so: object '/home/qeuser/libxphi/libxphi.so' from LD_PRELOAD cannot be preloaded (cannot open shared object file): ignored.
 buffer allocation   4.02019500732422      s

The AUSURF112 benchmark completed, it took very long time, and it reads

     PWSCF        :     9h56m CPU        0h36m WALL

   This run was terminated on:   2:25: 1  11Mar2017

=------------------------------------------------------------------------------=
   JOB DONE.
=------------------------------------------------------------------------------=
2
offload error: cannot unload library from the device 0 (error code 14)

On the host I can see one copy of pw.x is running, and on mic0 I can see that offload_main and coi_daemon are running by the micuser. But it is very slow.

So, is this offload error: cannot create buffer on device 0 (error code 14) related to the mpirun.sh script and the libxphi.so were not preloaded even it is present and the libxphi cannot be unload after completion???

With two Xeon E5-2683v3 CPU of 56 threads, it tooks only 9 mins and 15 seconds, but with this single instance offloaded to Xeon Phi, it tooks 10 hours?

I am running CentOS 7.1 + Intel MPSS 3.8.1 + Intel psxe 2017 update 1, and I have already made a symbolic link of the psxevars.sh to /etc/profile.d and I can use mpirun to pw.x on the host, but not offload to mic0 and mic1

Are these compatibility issues because the libxphi and mpirun.sh were written 2 years ago? How can these be fixed?

Thank you,

Rolly

0 Kudos

Link Copied

4 Replies
Rolly_N_
Beginner
‎03-27-2017 08:11 AM
1,327 Views

Hi all,

Thanks to Dr. Dahnken, developer of the libxphi library. He found the solution to my question.

https://github.com/cdahnken/libxphi

https://github.com/hfp/mpirun/blob/master/mpirun.py

In configuration with CentOS 7.1+ MPSS-3.8.1 + Intel PSXE 2017 update 1, a line have to be removed from the mpirun.py

runstring = "mpirun -bootstrap ssh"
if (None == args.mri): runstring = runstring \
                    + " -genv I_MPI_PIN_DOMAIN=auto" \
#                   + " -genv OFFLOAD_INIT=on_start" \
                    + " -genv MIC_USE_2MB_BUFFERS=2m" \
                    + " -genv MIC_ENV_PREFIX=" + micenv \
                    + " -genv " + micenv + "_KMP_AFFINITY=" + args.micaffinity + ",granularity=fine" \
                    + " -genv " + micenv + "_OMP_SCHEDULE=" + args.schedule \
                    + " -genv " + micenv + "_OMP_NUM_THREADS=" + str(max((mcores + min(0, args.reserved)) * 4, args.mthreads))

After removal of the line "+ " -genv OFFLOAD_INIT=on_start" \", it runs offloading to on 2x 7120P.

With these parameters, 2x 7120P complete the QE AUSURF112 benchmark in 17min56sec,

$ QE_MIC_BLOCKSIZE_M=2048
$ QE_MIC_BLOCKSIZE_N=2048
$ QE_MIC_BLOCKSIZE_K=512

By increasing K to 1024, 2x 7120P complete the QE AUSURF112 benchmark in 17min17sec,

I found 3 warning are given by the code:

(1) OMP: Info #256: KMP_PLACE_THREADS variable deprecated, please use KMP_HW_SUBSET instead.
(2) OMP: Warning #250: KMP_HW_SUBSET "o" offset designator deprecated, please use @ prefix for offset value.
(3) offload error: cannot unload library from the device 0 (error code 14)

and I am looking into these and I will report my findings.

Thanks for your attention.

Rolly

0 Kudos
Copy link
James_C_Intel2
Employee
‎03-27-2017 08:23 AM
1,327 Views

Re 1: KMP_PLACE_THREADS has been replaced with KMP_HW_SUBSET which better captures what the environmnet variable actually does.(The link below also points this out!)

Re 2: KMP_HW_SUBSET is documented in the compiler manual at https://software.intel.com/en-us/node/684224#DA65065D-8C30-4E3A-BA61-4AAE42DD505C (and elsewhere).

0 Kudos
Copy link
Rolly_N_
Beginner
‎03-28-2017 02:11 AM
1,327 Views

Hi James,

Thanks for the info.

I took a look at the link, I believe KMP_HW_SUBSET reads something like this, KMP_HW_SUBSET=socketsS[@offset],coresC[@offset],threadsT

The mpirun.py script generates something like this,

mpirun 
-bootstrap ssh 
-genv I_MPI_PIN_DOMAIN=auto 
-genv MIC_USE_2MB_BUFFERS=2m 
-genv MIC_ENV_PREFIX=MIC 
-genv MIC_KMP_AFFINITY=balanced,granularity=fine 
-genv MIC_OMP_SCHEDULE=static 
-genv MIC_OMP_NUM_THREADS=32 
-host node09 
-np 1 
-env KMP_AFFINITY=compact,1,granularity=fine 
-env OMP_SCHEDULE=static 
-env OMP_NUM_THREADS=4 
-env OFFLOAD_DEVICES=0 
-env MIC_KMP_PLACE_THREADS=8c,4t,0o 
-env MIC_OMP_SCHEDULE=static 
/home/qeuser/libxphi/xphilibwrapper.sh  
/home/qeuser/QE530-KNC-OL/espresso-5.3.0/bin/pw.x  < /home/qeuser/rolly/AUSURF112/ausurf.in : 
-host node09 
-np 1 
-env KMP_AFFINITY=compact,1,granularity=fine 
-env OMP_SCHEDULE=static 
-env OMP_NUM_THREADS=4 
-env OFFLOAD_DEVICES=0 
-env MIC_KMP_PLACE_THREADS=8c,4t,8o 
-env MIC_OMP_SCHEDULE=static 
/home/qeuser/libxphi/xphilibwrapper.sh  
/home/qeuser/QE530-KNC-OL/espresso-5.3.0/bin/pw.x  < /home/qeuser/rolly/AUSURF112/ausurf.in :
...
-host node09 
-np 1 
-env KMP_AFFINITY=compact,1,granularity=fine 
-env OMP_SCHEDULE=static 
-env OMP_NUM_THREADS=4 
-env OFFLOAD_DEVICES=1 
-env MIC_KMP_PLACE_THREADS=8c,4t,48o 
-env MIC_OMP_SCHEDULE=static 
/home/qeuser/libxphi/xphilibwrapper.sh  
/home/qeuser/QE530-KNC-OL/espresso-5.3.0/bin/pw.x  < /home/qeuser/rolly/AUSURF112/ausurf.in

So, is it correct to change environmental variable MIC_KMP_PLACE_THREADS to MIC_KMP_HW_SUBSET? Shall it looks like this?

-host node09 
-np 1 
-env KMP_AFFINITY=compact,1,granularity=fine 
-env OMP_SCHEDULE=static 
-env OMP_NUM_THREADS=4 
-env OFFLOAD_DEVICES=0 
-env MIC_KMP_HW_SUBSET=0s,8c@0,4t 
-env MIC_OMP_SCHEDULE=static 
/home/qeuser/libxphi/xphilibwrapper.sh  
/home/qeuser/QE530-KNC-OL/espresso-5.3.0/bin/pw.x  < /home/qeuser/rolly/AUSURF112/ausurf.in :
-host node09 
-np 1 
-env KMP_AFFINITY=compact,1,granularity=fine 
-env OMP_SCHEDULE=static 
-env OMP_NUM_THREADS=4 
-env OFFLOAD_DEVICES=0 
-env MIC_KMP_HW_SUBSET=0s,8c@8,4t 
-env MIC_OMP_SCHEDULE=static 
/home/qeuser/libxphi/xphilibwrapper.sh  
/home/qeuser/QE530-KNC-OL/espresso-5.3.0/bin/pw.x  < /home/qeuser/rolly/AUSURF112/ausurf.in :
...
-host node09 
-np 1 
-env KMP_AFFINITY=compact,1,granularity=fine 
-env OMP_SCHEDULE=static 
-env OMP_NUM_THREADS=4 
-env OFFLOAD_DEVICES=1 
-env MIC_KMP_HW_SUBSET=0s@1,8c@48,4t 
-env MIC_OMP_SCHEDULE=static 
/home/qeuser/libxphi/xphilibwrapper.sh  
/home/qeuser/QE530-KNC-OL/espresso-5.3.0/bin/pw.x  < /home/qeuser/rolly/AUSURF112/ausurf.in :

The MIC_KMP_PLACE_THREADS also assign different thread numbers to different core via the "o" offset designator, but how can I do the same with MIC_KMP_HW_SUBSET? By the way, is MIC_KMP_HW_SUBSET the correct environmental variable?

Thanks,

Rolly

0 Kudos
Copy link
Rolly_N_
Beginner
‎03-29-2017 07:37 PM
1,327 Views

Hi all,

I am able to remove the first two warnings by modifying the mpirun.py. Can anyone check if these are correct? Thank you!

#! /usr/bin/env python
###############################################################################
## Copyright (c) 2014-2015, Intel Corporation                                ##
## All rights reserved.                                                      ##
##                                                                           ##
## Redistribution and use in source and binary forms, with or without        ##
## modification, are permitted provided that the following conditions        ##
## are met:                                                                  ##
## 1. Redistributions of source code must retain the above copyright         ##
##    notice, this list of conditions and the following disclaimer.          ##
## 2. Redistributions in binary form must reproduce the above copyright      ##
##    notice, this list of conditions and the following disclaimer in the    ##
##    documentation and/or other materials provided with the distribution.   ##
## 3. Neither the name of the copyright holder nor the names of its          ##
##    contributors may be used to endorse or promote products derived        ##
##    from this software without specific prior written permission.          ##
##                                                                           ##
## THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS       ##
## "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT         ##
## LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR     ##
## A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT      ##
## HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,    ##
## SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED  ##
## TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR    ##
## PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF    ##
## LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING      ##
## NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS        ##
## SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.              ##
###############################################################################
## Hans Pabst, Christopher Dahnken, Intel Corporation                        ##
###############################################################################
import platform
import sys
import os

try:
    import argparse
except ImportError, e:
    print 'Failed to import the argparse module: ', e
    print 'Please follow the troubleshooting instructions at:'
    print 'https://github.com/hfp/mpirun#troubleshooting'


def micshift(i, rcores, ncores, nthreads):
    return str(max(ncores + min(0, rcores), nthreads)) + "c@" + str(i * ncores) + "," + str(nthreads) + "t"  


nodename = platform.node().split(".")[0]
micenv = "MIC"

parser = argparse.ArgumentParser()
parser.add_argument("-n", "--nodelist", help="list of comma separated node names", default=["localhost", nodename]["" != nodename])
parser.add_argument("-p", "--cpuprocs", help="number of processes per socket (host)", type=int, default=1)
parser.add_argument("-q", "--micprocs", help="number of processes per mic (native)", type=int, default=1)
parser.add_argument("-s", "--nsockets", help="number of sockets per node", type=int, default=1)
parser.add_argument("-d", "--ndevices", help="number of devices per node", type=int, default=0)
parser.add_argument("-e", "--cpucores", help="number of CPU cores per socket", type=int, default=1)
parser.add_argument("-t", "--nthreads", help="number of CPU threads per core", type=int, default=2)
parser.add_argument("-m", "--miccores", help="number of MIC cores per device", type=int, default=57)
parser.add_argument("-r", "--reserved", help="number of MIC cores reserved", type=int, default=sys.maxint)
parser.add_argument("-u", "--mthreads", help="number of MIC threads per core", type=int, default=4)
parser.add_argument("-a", "--cpuaffinity", help="affinity (CPU) e.g., compact", default="compact")
parser.add_argument("-b", "--micaffinity", help="affinity (MIC) e.g., balanced", default="balanced")
parser.add_argument("-c", "--schedule", help="schedule, e.g. dynamic", default="static")
parser.add_argument("-w", "--wrapper", help="wrapper")
parser.add_argument("-g", "--debugger", help="debugger")
parser.add_argument("-i", "--inputfile", help="inputfile (<)")
parser.add_argument("-0", "--hr0", help="executable (rank-0)")
parser.add_argument("-x", "--hri", help="executable (host)")
parser.add_argument("-y", "--mri", help="executable (mic)")
parser.add_argument("-z", "--micpre", help="prefixed mic name", action="store_true")
parser.add_argument("-v", "--dryrun", help="dryrun", action="store_true")
args, unknown = parser.parse_known_args()

if (None != args.inputfile):
    arguments = " ".join(unknown) + " < " + args.inputfile
else:
    arguments = " ".join(unknown)
if ("" != arguments): arguments = " " + arguments

if (None != args.wrapper):
    wrapper = args.wrapper + [" ", " " + str(args.debugger) + " -ex run --args"][None != args.debugger]
else:
    wrapper = ["", str(args.debugger) + " -ex run --args"][None != args.debugger]
if ("" != wrapper): wrapper = wrapper + " "

if (1 < args.nsockets): args.cpuaffinity = args.cpuaffinity + ",1"
if (1 < args.nthreads): args.cpuaffinity = args.cpuaffinity + ",granularity=fine"

if (None == args.mri):
    if (sys.maxint == args.reserved): args.reserved = 1
    if (0 < args.ndevices):
        cpusockets = min(args.nsockets, args.ndevices)
    else:
        cpusockets = args.nsockets
    nparts = args.cpuprocs
else:
    if (sys.maxint == args.reserved): args.reserved = 0
    cpusockets = args.nsockets
    nparts = args.micprocs

cputhreads = cpusockets * args.cpucores * args.nthreads
mcores = int(0 < nparts) * max((args.miccores - max(args.reserved, 0)), nparts) / max(nparts, 1)
pthreads = int(0 < (args.cpuprocs * cpusockets)) * cputhreads / max(args.cpuprocs * cpusockets, 1)
remainder = cputhreads - pthreads * args.cpuprocs * cpusockets


runstring = "mpirun -bootstrap ssh"
if (None == args.mri): runstring = runstring \
                    + " -genv I_MPI_PIN_DOMAIN=auto" \
		    + " -genv MIC_USE_2MB_BUFFERS=2m" \
                    + " -genv MIC_ENV_PREFIX=" + micenv \
                    + " -genv " + micenv + "_KMP_AFFINITY=" + args.micaffinity + ",granularity=fine" \
                    + " -genv " + micenv + "_OMP_SCHEDULE=" + args.schedule \
                    + " -genv " + micenv + "_OMP_NUM_THREADS=" + str(max((mcores + min(0, args.reserved)) * 4, args.mthreads))
else: runstring = runstring \
                    + " -genv I_MPI_MIC=1"
if (None != args.hr0): runstring = runstring \
                    + " -host " + args.nodelist.split(",")[0] + " -np 1" \
                    + " -env I_MPI_PIN_DOMAIN=auto" \
                    + " -env KMP_AFFINITY=" + args.cpuaffinity \
                    + " -env OMP_NUM_THREADS=" + str(args.nthreads) \
                    + " " + wrapper + args.hr0 + arguments \
                    + " :"
for n in args.nodelist.split(","):
    for s in range(0, cpusockets):
        for p in range(0, args.cpuprocs):
            runstring = runstring \
                    + " -host " + n + " -np 1" \
                    + " -env KMP_AFFINITY=" + args.cpuaffinity \
                    + " -env OMP_SCHEDULE=" + args.schedule \
                    + " -env OMP_NUM_THREADS=" + str(pthreads - [0, args.nthreads][None != args.hr0 and 0 == s and 0 == p])
            if (None == args.mri): runstring = runstring \
                    + " -env OFFLOAD_DEVICES=" + str(s) \
                    + " -env " + micenv + "_KMP_HW_SUBSET=" + str(s) + "s," + micshift(p, args.reserved, mcores, args.mthreads) \
                    + " -env " + micenv + "_OMP_SCHEDULE=" + args.schedule
            else: runstring = runstring \
                    + " -env I_MPI_PIN_DOMAIN=auto"
            if (None != args.hri): runstring = runstring \
                    + " " + wrapper + args.hri + arguments \
                    + " :"
    if (None != args.mri):
        for d in range(0, args.ndevices):
            for m in range(0, args.micprocs):
                runstring = runstring \
                    + " -host " + ["", n + "-"][int(args.micpre)] + "mic" + str(d) + " -np 1" \
                    + " -env LD_LIBRARY_PATH=$MIC_LD_LIBRARY_PATH" \
                    + " -env I_MPI_PIN=off" \
                    + " -env KMP_HW_SUBSET=" + str(d) + "s," + micshift(m, args.reserved, mcores, args.mthreads) \
                    + " -env KMP_AFFINITY=" + args.micaffinity + ",granularity=fine" \
                    + " -env OMP_SCHEDULE=" + args.schedule \
                    + " -env OMP_NUM_THREADS=" + str(max(mcores + min(0, args.reserved), args.mthreads) * 4) \
                    + " " + args.mri + arguments \
                    + " :"
        if (None != args.hri and 0 < remainder and 0 < cputhreads and 0 < args.cpuprocs): runstring = runstring \
                    + " -host " + n + " -np 1" \
                    + " -env I_MPI_PIN_DOMAIN=auto" \
                    + " -env KMP_AFFINITY=" + args.cpuaffinity \
                    + " -env OMP_SCHEDULE=" + args.schedule \
                    + " -env OMP_NUM_THREADS=" + str(remainder) \
                    + " " + wrapper + args.hri + arguments \
                    + " :"

runstring = runstring[0:len(runstring)-2]
print runstring
print

result = 0
if (False == args.dryrun):
    result = os.system(runstring)

sys.exit([0, 1][0 != result])

But for the 3rd warning, (3) offload error: cannot unload library from the device 0 (error code 14), can anyone help?

Thanks,

Rolly

0 Kudos
Copy link
Reply

Community support is provided Monday to Friday. Other contact methods are available here.

Intel does not verify all solutions, including but not limited to any file transfers that may appear in this community. Accordingly, Intel disclaims all express and implied warranties, including without limitation, the implied warranties of merchantability, fitness for a particular purpose, and non-infringement, as well as any warranty arising from course of performance, course of dealing, or usage in trade.

For more complete information about compiler optimizations, see our Optimization Notice.