July 2014 - discuss - mailman3.cels.anl.gov

mpicc shell error of some kind
by Brian Van Straalen 02 Jul '14

02 Jul '14

an interesting error while using mpicc with mpich-3.1.1 >which mpicc /opt/pkg/mpich-3.1.1/bin/mpicc > mpicc -compile-info gcc -I/opt/pkg/mpich-3.1.1/include -L/opt/pkg/mpich-3.1.1/lib -Wl,-rpath -Wl,/opt/pkg/mpich-3.1.1/lib -Wl,--enable-new-dtags -lmpi -lrt -lpthread >mpicc -DHAVE_CONFIG_H -I. -I.. -I/opt/pkg/lrds-0.5/include -Wall -g -I../include -D_LARGEFILE64_SOURCE -O0 -g -I/opt/pkg/lrds-0.5/include -MT mcs-mutex.lo -MD -MP -MF .deps/mcs-mutex.Tpo -c mcs-mutex.c -fPIC -v -DPIC -o .libs/mcs-mutex.o ///// bunch of stuff about gcc 4.6.3 since I have the -v flag added here, but then you notice this result COLLECT_GCC_OPTIONS='-D' 'HAVE_CONFIG_H' '-I' '.' '-I' '..' '-I' '/opt/pkg/lrds-0.5/include' '-Wall' '-g' '-I' '../include' '-D' '_LARGEFILE64_SOURCE' '-O0' '-g' '-I' '/opt/pkg/lrds-0.5/include' '-MT' 'mcs-mutex.lo' '-MD' '-MP' '-MF' '.deps/mcs-mutex.Tpo' '-c' '-fPIC' '-v' '-D' 'PIC' '-o' '.libs/mcs-mutex.o' '-I' '/usr/lib/openmpi/include' '-I' '/usr/lib/openmpi/include/openmpi' '-pthread' '-mtune=generic' '-march=x86-64' what has gone missing is every flag that mpicc thought is was giving gcc. no mpich-3.1.1/include, no flags, nothing. it has all been eaten. as a result I end up picking up an mpi.h sitting in /usr/local/include, and get errors about not finding any one-sided MPI function declarations. when I substitute the compile-info string in for mpicc things work correctly. why is mpicc failing to send it’s flags to gcc correctly? Brian Van Straalen Lawrence Berkeley Lab

1 0

Re: [mpich-discuss] Possible bug MPICH2 1.0.6
by mark 02 Jul '14

02 Jul '14

???? 02/07/2014 11:02 ??, ?/? mark ??????: > Hello. > In a program i am writing, i have an array which size is equal to the > number of nodes in my cluster. > I made this array to keep an eye of the objects that each node has in > his control and when i use it with the actual ranks of the nodes as > positions ( *array[node_rank]=objects;*) it returns me an error and > stops the execution: > > *rank 0 in job 10 Calliope_49755 caused collective abort of all ranks** > ** exit status of rank 0: killed by signal 11 * > > However, i had freely used the variable /id/ or /node_rank/ to point > out a specific action for a specific node since now and never had > problems. > > If instead of using the ranks node as a pointer for the array, i use > an integer i.e. 5, the program runs fine, but even now, 1 out of 7 > executions, it returns me an error and stops. The action i am calling > that keeps returning me the error is a simple abstraction > (*array[id]--;*). > > Any ideas? > > > PS. The array is 1d and is created by allocating memory, like this : > *int *array = malloc(processes * sizeof(int));* I forgot to mention how i compiled and executed the program : mpich -o prog prog.c -lm mpiexec -n nodes `pwd`/prog

1 0

Possible bug MPICH2 1.0.6
by mark 02 Jul '14

02 Jul '14

Hello. In a program i am writing, i have an array which size is equal to the number of nodes in my cluster. I made this array to keep an eye of the objects that each node has in his control and when i use it with the actual ranks of the nodes as positions ( *array[node_rank]=objects;*) it returns me an error and stops the execution: *rank 0 in job 10 Calliope_49755 caused collective abort of all ranks** ** exit status of rank 0: killed by signal 11 * However, i had freely used the variable /id/ or /node_rank/ to point out a specific action for a specific node since now and never had problems. If instead of using the ranks node as a pointer for the array, i use an integer i.e. 5, the program runs fine, but even now, 1 out of 7 executions, it returns me an error and stops. The action i am calling that keeps returning me the error is a simple abstraction (*array[id]--;*). Any ideas? PS. The array is 1d and is created by allocating memory, like this : *int *array = malloc(processes * sizeof(int));*

1 0

Re: [mpich-discuss] Time distribution of Intel MPI Benchmark (PingPong)
by "Antonio J. Peña" 02 Jul '14

02 Jul '14

Unfortunately the default communication channel doesn't have that capability. You'll have to recompile MPICH configuring with --with-device=ch3:sock Antonio On 07/02/2014 01:54 AM, Klaus Philip wrote: > > Dear Antonio, > > Thank you for your fast answering. > > Do you know how to disable this "active polling" ? > > Greetings, > > Philip > > *Von:*[email protected] [mailto:[email protected]] *Im > Auftrag von *"Antonio J. Peï¿½a" > *Gesendet:* Dienstag, 1. Juli 2014 16:30 > *An:* discuss(a)mpich.org > *Betreff:* Re: [mpich-discuss] Time distribution of Intel MPI > Benchmark (PingPong) > > > Dear Philip, > > Both functions are part of MPICH. MPICH uses intentionally active > polling to minimize the latency on communications. In short, you get > faster communications by repeatedly checking for communication > completion during waiting times. So, that's expected. > > Antonio > > > On 07/01/2014 09:16 AM, Klaus Philip wrote: > > Hi, > > I tested mpich2 3.1.1 with the Intel MPI Benchmark, more precisely > the "PingPong"-Benchmark. > > It uses the MPI_Send and MPI_Recv function to send one message > from /Process 1/ to /Process 2/ and back. > > While the benchmark was running, I executed the Sysprofiler > (sysprof) to check the time distribution (Picture below). > > My question is now, if the two marked Processes > (*MPID_nem_lmt_shm_progress* and *poll*) are parts of MPICH or the > Intel-PingPong-Benchmark, because they are both consuming much > time (49,03% and 19,99%). > > Greetings, > > Philip > > > > > _______________________________________________ > > discuss mailing listdiscuss(a)mpich.org <mailto:[email protected]> > > To manage subscription options or unsubscribe: > > https://lists.mpich.org/mailman/listinfo/discuss > > > > > -- > Antonio J. Peï¿½a > Postdoctoral Appointee > Mathematics and Computer Science Division > Argonne National Laboratory > 9700 South Cass Avenue, Bldg. 240, Of. 3148 > Argonne, IL 60439-4847 > apenya(a)mcs.anl.gov <mailto:[email protected]> > www.mcs.anl.gov/~apenya <http://www.mcs.anl.gov/%7Eapenya> > > > _______________________________________________ > discuss mailing list discuss(a)mpich.org > To manage subscription options or unsubscribe: > https://lists.mpich.org/mailman/listinfo/discuss -- Antonio J. Peï¿½a Postdoctoral Appointee Mathematics and Computer Science Division Argonne National Laboratory 9700 South Cass Avenue, Bldg. 240, Of. 3148 Argonne, IL 60439-4847 apenya(a)mcs.anl.gov www.mcs.anl.gov/~apenya

1 0

Re: [mpich-discuss] Time distribution of Intel MPI Benchmark (PingPong)
by Klaus Philip 02 Jul '14

02 Jul '14

Dear Antonio, Thank you for your fast answering. Do you know how to disable this "active polling" ? Greetings, Philip Von: discuss-bounces(a)mpich.org [mailto:[email protected]] Im Auftrag von "Antonio J. Peña" Gesendet: Dienstag, 1. Juli 2014 16:30 An: discuss(a)mpich.org Betreff: Re: [mpich-discuss] Time distribution of Intel MPI Benchmark (PingPong) Dear Philip, Both functions are part of MPICH. MPICH uses intentionally active polling to minimize the latency on communications. In short, you get faster communications by repeatedly checking for communication completion during waiting times. So, that's expected. Antonio On 07/01/2014 09:16 AM, Klaus Philip wrote: Hi, I tested mpich2 3.1.1 with the Intel MPI Benchmark, more precisely the "PingPong"-Benchmark. It uses the MPI_Send and MPI_Recv function to send one message from Process 1 to Process 2 and back. While the benchmark was running, I executed the Sysprofiler (sysprof) to check the time distribution (Picture below). My question is now, if the two marked Processes (MPID_nem_lmt_shm_progress and poll) are parts of MPICH or the Intel-PingPong-Benchmark, because they are both consuming much time (49,03% and 19,99%). [cid:[email protected]] Greetings, Philip _______________________________________________ discuss mailing list discuss(a)mpich.org<mailto:[email protected]> To manage subscription options or unsubscribe: https://lists.mpich.org/mailman/listinfo/discuss -- Antonio J. Peña Postdoctoral Appointee Mathematics and Computer Science Division Argonne National Laboratory 9700 South Cass Avenue, Bldg. 240, Of. 3148 Argonne, IL 60439-4847 apenya(a)mcs.anl.gov<mailto:[email protected]> www.mcs.anl.gov/~apenya<http://www.mcs.anl.gov/~apenya>

1 0

Fwd: EuroMPI/ASIA 2014 Call for Participation
by Rajeev Thakur 01 Jul '14

01 Jul '14

EuroMPI/ASIA 2014 Call for participation EuroMPI/ASIA 2014 in-cooperation status with ACM and SIGHPC in Kyoto, Japan, 9th - 12th September, 2014. The prime annual meeting for researchers, developers, and students in message-passing parallel computing with MPI and related paradigms. **** Deadline of early registration is July 31, 2014 ***** www.eurompi2014.org The conference will feature 19 strong technical paper presentations, 3 invited talks, 2 tutorials, 2 workshops, posters ,and exhibitions. The list of accepted papers is attached in this CFP. The detailed conference information is being incrementally updated in www.eurompi2014.org. Please watch the URL to get the latest information. TUTORIALS ------------------------------- - Advanced MPI: New Features of MPI-3 Torsten Hoefler, ETH, Switzerland - Practical Parallel Application Performance Engineering Marc-Andre Hermanns, German Research School for Simulation Sciences, Germany Allen Malony, University of Oregon Matthias Weber, TU Dresden, Germany INVITED TALKS ------------------------------- - Enabling Scientific Discovery Through High Performance Computing and Extreme Data Science with the Cori System Katie Antypas, National Energy Research Scientific Computing Center (NERSC), USA - Large Scale System Design and Application for Tianhe-2 Yutong Lu, National University of Defense Technology (NUDT), China - NESUS: Looking for sustainability in ultrascale computing systems Jesus Carretero, Universidad Carlos III de Madrid, Spain WORKSHOPS ------------------------------- - Challenges in Data-Centric Computing https://www.hlrs.de/index.php?id=2032 - ESAA, Workshop on Enhancing Parallel Scientific Applications with Accelerated HPC http://www.arcos.inf.uc3m.es/~esaa2014/ IMPORTANT DATES ------------------------------- - Early Registration Deadline: July 31, 2014 - Visa support Deadline: August 8, 2014 - Pre-Registration Deadline: August 25, 2014 - Tutorials: September 9th, 2014 - Conference: September 10th-12th, 2014 ACCEPTED PAPERS ------------------------------- - MPI Collectives and Datatypes for hierarchical All-to-all Communication Jesper Larsson Traff and Antoine Rougier - Optimal MPI datatype normalization for vector and index-block types Jesper Larsson Traff - Zero-copy, hierarchical Gather is not possible with MPI Datatypes and Collectives Jesper Larsson Tr«£ff and Antoine Rougier - GPU-Aware Intranode MPI_Allreduce Iman Faraji and Ahmad Afsahi - Toward Local Failure Local Recovery Resilience Model using MPI-ULFM Keita Teranishi and Michael Heroux - Evaluating User-Level Fault Tolerance for MPI Applications Ignacio Laguna, David F. Richards, Todd Gamblin, Martin Schulz and Bronis R. de Supinski - Comparing, Contrasting, Generalizing, and Integrating Two Current Designs for Fault-Tolerant MPI Amin Hassani, Anthony Skjellum and Ron Brightwell - Implementing the MPI-3.0 Fortran 2008 Binding Junchao Zhang, Bill Long, Kenneth Raffenetti and Pavan Balaji - Scalable MPI3 RMA on the Blue Gene/Q Supercomputer Sameer Kumar and Michael Blocksome - Intra-Epoch Message Scheduling To Exploit Unused or Residual Overlapping Potential Judicael A. Zounmevo and Ahmad Afsahi - PMI Extensions for Scalable MPI Startup Sourav Chakraborty, Hari Subramoni, Jonathan Perkins, Adam Moody, Mark Arnold and Dhabaleswar Panda - Exploring the Capabilities of the New MPI_T Interface Tanzima Islam, Kathryn Mohror and Martin Schulz - Understanding the Memory-Utilization of MPI Libraries: Challenges and Designs in Implementing the MPI_T Interface Raghunath Rajachandrasekar, Jonathan Perkins, Khaled Hamidouche, Mark Arnold and Dhabaleswar K. Panda - Catching Idlers with Ease: A Lightweight Wait-State Profiler for MPI Programs Guoyong Mao, David Boehme, Marc-Andre Hermanns, Markus Geimer, Daniel Lorenz and Felix Wolf - Distributed Behavioral Cartography of Timed Automata Etienne Andre, Camille Coti and Sami Evangelista - Reproducible MPI Micro-Benchmarking Isn't As Easy As You Think Sascha Hunold, Alexandra Carpen-Amarie and Jesper Larsson Tr«£ff - Exploring the effect of noise on the performance benefit of non-blocking MPI_Allreduce Patrick Widener, Kurt Ferreira and Scott Levy - A Portable Petascale Framework for Efficient Particle Methods with Custom Interactions Andreas Schafer and Dietmar Fey

1 0

Re: [mpich-discuss] how to disable fortran during mpich installation?
by Kenneth Raffenetti 01 Jul '14

01 Jul '14

If you are building mpich-3.1.1, the flag to disable fortran is "--disable-fortran". On 07/01/2014 09:55 AM, Rintu Nath wrote: > how to disable fortran during mpich installation? My installation is not > complete as I am unable to disable fortran. Installing in Ubuntu 12.01 > > > > > _______________________________________________ > discuss mailing list discuss(a)mpich.org > To manage subscription options or unsubscribe: > https://lists.mpich.org/mailman/listinfo/discuss >

1 0

how to disable fortran during mpich installation?
by Rintu Nath 01 Jul '14

01 Jul '14

how to disable fortran during mpich installation? My installation is not complete as I am unable to disable fortran. Installing in Ubuntu 12.01

1 0

Re: [mpich-discuss] Time distribution of Intel MPI Benchmark (PingPong)
by "Antonio J. Peña" 01 Jul '14

01 Jul '14

Dear Philip, Both functions are part of MPICH. MPICH uses intentionally active polling to minimize the latency on communications. In short, you get faster communications by repeatedly checking for communication completion during waiting times. So, that's expected. Antonio On 07/01/2014 09:16 AM, Klaus Philip wrote: > > Hi, > > I tested mpich2 3.1.1 with the Intel MPI Benchmark, more precisely the > "PingPong"-Benchmark. > > It uses the MPI_Send and MPI_Recv function to send one message from > /Process 1/ to /Process 2/ and back. > > While the benchmark was running, I executed the Sysprofiler (sysprof) > to check the time distribution (Picture below). > > My question is now, if the two marked Processes > (*MPID_nem_lmt_shm_progress* and *poll*) are parts of MPICH or the > Intel-PingPong-Benchmark, because they are both consuming much time > (49,03% and 19,99%). > > Greetings, > > Philip > > > > _______________________________________________ > discuss mailing list discuss(a)mpich.org > To manage subscription options or unsubscribe: > https://lists.mpich.org/mailman/listinfo/discuss -- Antonio J. Peï¿½a Postdoctoral Appointee Mathematics and Computer Science Division Argonne National Laboratory 9700 South Cass Avenue, Bldg. 240, Of. 3148 Argonne, IL 60439-4847 apenya(a)mcs.anl.gov www.mcs.anl.gov/~apenya

1 0

Time distribution of Intel MPI Benchmark (PingPong)
by Klaus Philip 01 Jul '14

01 Jul '14

Hi, I tested mpich2 3.1.1 with the Intel MPI Benchmark, more precisely the "PingPong"-Benchmark. It uses the MPI_Send and MPI_Recv function to send one message from Process 1 to Process 2 and back. While the benchmark was running, I executed the Sysprofiler (sysprof) to check the time distribution (Picture below). My question is now, if the two marked Processes (MPID_nem_lmt_shm_progress and poll) are parts of MPICH or the Intel-PingPong-Benchmark, because they are both consuming much time (49,03% and 19,99%). [cid:[email protected]] Greetings, Philip

1 0