December 2017 - discuss - mailman3.cels.anl.gov

Re: [mpich-discuss] how to run jobs using IPv6?
by Guo, Yanfei 06 Dec '17

06 Dec '17

I have created an ticket to track it. https://github.com/pmodels/mpich/issues/2889 Yanfei Guo Assistant Computer Scientist MCS Division, ANL On 12/6/17, 11:52 AM, "Shuwei Zhao" <shuweizhao1991(a)gmail.com> wrote: Are you going to use getaddrinfo instead to deprecate gethostbyname? Thanks On Wed, 6 Dec 2017 at 11:50 Guo, Yanfei <yguo(a)anl.gov> wrote: Hi Shuwei, The communication over IPv6 network is supported as long as the lower-level network library supports it. However, MPICH uses gethostbyname to translate hostnames to addresses in order to initialize the link which is not working on pure IPv6. Yanfei Guo Assistant Computer Scientist MCS Division, ANL On 12/6/17, 11:46 AM, "Shuwei Zhao" <shuweizhao1991(a)gmail.com> wrote: Hi Yanfei， I ask this question months ago，but I was told IPv6 is supported in mpich 3.*. So it's not supported? Thanks Shuwei On Wed, 6 Dec 2017 at 11:24 Guo, Yanfei <yguo(a)anl.gov> wrote: Hi Carlos, MPICH currently uses gethostbyname which does not support IPv6, therefore it won’t work on a pure IPv6 network. Best, Yanfei Guo Assistant Computer Scientist MCS Division, ANL On 11/22/17, 2:23 PM, "Carlos" <carlos(a)inf.ufpr.br> wrote: Is it possible to use mpich (3.3) with IPv6-only? We tried it here and got Fatal error in MPI_Init: Other MPI error, error stack: MPIR_Init_thread(506)..............: MPID_Init(211).....................: channel initialization failed MPIDI_CH3_Init(89).................: MPID_nem_init(324).................: MPID_nem_tcp_init(178).............: MPID_nem_tcp_get_business_card(425): ckpt_restart(384)..................: gethostbyname failed, zuratkul (errno 4) ************* This shows that mpich doesn't run without IPv4, which is our case... Is there a way to make it work in this situation? _______________________________________________ discuss mailing list discuss(a)mpich.org To manage subscription options or unsubscribe: https://lists.mpich.org/mailman/listinfo/discuss _______________________________________________ discuss mailing list discuss(a)mpich.org To manage subscription options or unsubscribe: https://lists.mpich.org/mailman/listinfo/discuss _______________________________________________ discuss mailing list discuss(a)mpich.org To manage subscription options or unsubscribe: https://lists.mpich.org/mailman/listinfo/discuss _______________________________________________ discuss mailing list discuss(a)mpich.org To manage subscription options or unsubscribe: https://lists.mpich.org/mailman/listinfo/discuss

1 0

Re: [mpich-discuss] how to run jobs using IPv6?
by Shuwei Zhao 06 Dec '17

06 Dec '17

Are you going to use getaddrinfo instead to deprecate gethostbyname? Thanks On Wed, 6 Dec 2017 at 11:50 Guo, Yanfei <yguo(a)anl.gov> wrote: > Hi Shuwei, > > The communication over IPv6 network is supported as long as the > lower-level network library supports it. However, MPICH uses gethostbyname > to translate hostnames to addresses in order to initialize the link which > is not working on pure IPv6. > > Yanfei Guo > Assistant Computer Scientist > MCS Division, ANL > > > > On 12/6/17, 11:46 AM, "Shuwei Zhao" <shuweizhao1991(a)gmail.com> wrote: > > Hi Yanfei， > > > I ask this question months ago，but I was told IPv6 is supported in mpich > 3.*. So it's not supported? > > > Thanks > Shuwei > > On Wed, 6 Dec 2017 at 11:24 Guo, Yanfei <yguo(a)anl.gov> wrote: > > > Hi Carlos, > > MPICH currently uses gethostbyname which does not support IPv6, therefore > it won’t work on a pure IPv6 network. > > Best, > > Yanfei Guo > Assistant Computer Scientist > MCS Division, ANL > > > > On 11/22/17, 2:23 PM, "Carlos" <carlos(a)inf.ufpr.br> wrote: > > Is it possible to use mpich (3.3) with IPv6-only? We tried it here and got > > Fatal error in MPI_Init: Other MPI error, error stack: > MPIR_Init_thread(506)..............: > MPID_Init(211).....................: channel initialization failed > MPIDI_CH3_Init(89).................: > MPID_nem_init(324).................: > MPID_nem_tcp_init(178).............: > MPID_nem_tcp_get_business_card(425): > ckpt_restart(384)..................: gethostbyname failed, zuratkul (errno > 4) > ************* > > This shows that mpich doesn't run without IPv4, which is our case... Is > there a > way to make it work in this situation? > _______________________________________________ > discuss mailing list discuss(a)mpich.org > To manage subscription options or unsubscribe: > https://lists.mpich.org/mailman/listinfo/discuss > > > _______________________________________________ > discuss mailing list discuss(a)mpich.org > To manage subscription options or unsubscribe: > https://lists.mpich.org/mailman/listinfo/discuss > > > > > > _______________________________________________ > discuss mailing list discuss(a)mpich.org > To manage subscription options or unsubscribe: > https://lists.mpich.org/mailman/listinfo/discuss _______________________________________________ discuss mailing list discuss(a)mpich.org To manage subscription options or unsubscribe: https://lists.mpich.org/mailman/listinfo/discuss

1 0

Re: [mpich-discuss] how to run jobs using IPv6?
by Guo, Yanfei 06 Dec '17

06 Dec '17

Hi Shuwei, The communication over IPv6 network is supported as long as the lower-level network library supports it. However, MPICH uses gethostbyname to translate hostnames to addresses in order to initialize the link which is not working on pure IPv6. Yanfei Guo Assistant Computer Scientist MCS Division, ANL On 12/6/17, 11:46 AM, "Shuwei Zhao" <shuweizhao1991(a)gmail.com> wrote: Hi Yanfei， I ask this question months ago，but I was told IPv6 is supported in mpich 3.*. So it's not supported? Thanks Shuwei On Wed, 6 Dec 2017 at 11:24 Guo, Yanfei <yguo(a)anl.gov> wrote: Hi Carlos, MPICH currently uses gethostbyname which does not support IPv6, therefore it won’t work on a pure IPv6 network. Best, Yanfei Guo Assistant Computer Scientist MCS Division, ANL On 11/22/17, 2:23 PM, "Carlos" <carlos(a)inf.ufpr.br> wrote: Is it possible to use mpich (3.3) with IPv6-only? We tried it here and got Fatal error in MPI_Init: Other MPI error, error stack: MPIR_Init_thread(506)..............: MPID_Init(211).....................: channel initialization failed MPIDI_CH3_Init(89).................: MPID_nem_init(324).................: MPID_nem_tcp_init(178).............: MPID_nem_tcp_get_business_card(425): ckpt_restart(384)..................: gethostbyname failed, zuratkul (errno 4) ************* This shows that mpich doesn't run without IPv4, which is our case... Is there a way to make it work in this situation? _______________________________________________ discuss mailing list discuss(a)mpich.org To manage subscription options or unsubscribe: https://lists.mpich.org/mailman/listinfo/discuss _______________________________________________ discuss mailing list discuss(a)mpich.org To manage subscription options or unsubscribe: https://lists.mpich.org/mailman/listinfo/discuss _______________________________________________ discuss mailing list discuss(a)mpich.org To manage subscription options or unsubscribe: https://lists.mpich.org/mailman/listinfo/discuss

1 0

Re: [mpich-discuss] how to run jobs using IPv6?
by Shuwei Zhao 06 Dec '17

06 Dec '17

Hi Yanfei， I ask this question months ago，but I was told IPv6 is supported in mpich 3.*. So it's not supported? Thanks Shuwei On Wed, 6 Dec 2017 at 11:24 Guo, Yanfei <yguo(a)anl.gov> wrote: > Hi Carlos, > > MPICH currently uses gethostbyname which does not support IPv6, therefore > it won’t work on a pure IPv6 network. > > Best, > > Yanfei Guo > Assistant Computer Scientist > MCS Division, ANL > > > > On 11/22/17, 2:23 PM, "Carlos" <carlos(a)inf.ufpr.br> wrote: > > Is it possible to use mpich (3.3) with IPv6-only? We tried it here and got > > Fatal error in MPI_Init: Other MPI error, error stack: > MPIR_Init_thread(506)..............: > MPID_Init(211).....................: channel initialization failed > MPIDI_CH3_Init(89).................: > MPID_nem_init(324).................: > MPID_nem_tcp_init(178).............: > MPID_nem_tcp_get_business_card(425): > ckpt_restart(384)..................: gethostbyname failed, zuratkul (errno > 4) > ************* > > This shows that mpich doesn't run without IPv4, which is our case... Is > there a > way to make it work in this situation? > _______________________________________________ > discuss mailing list discuss(a)mpich.org > To manage subscription options or unsubscribe: > https://lists.mpich.org/mailman/listinfo/discuss > > > _______________________________________________ > discuss mailing list discuss(a)mpich.org > To manage subscription options or unsubscribe: > https://lists.mpich.org/mailman/listinfo/discuss _______________________________________________ discuss mailing list discuss(a)mpich.org To manage subscription options or unsubscribe: https://lists.mpich.org/mailman/listinfo/discuss

1 0

Re: [mpich-discuss] how to run jobs using IPv6?
by Guo, Yanfei 06 Dec '17

06 Dec '17

Hi Carlos, MPICH currently uses gethostbyname which does not support IPv6, therefore it won’t work on a pure IPv6 network. Best, Yanfei Guo Assistant Computer Scientist MCS Division, ANL On 11/22/17, 2:23 PM, "Carlos" <carlos(a)inf.ufpr.br> wrote: Is it possible to use mpich (3.3) with IPv6-only? We tried it here and got Fatal error in MPI_Init: Other MPI error, error stack: MPIR_Init_thread(506)..............: MPID_Init(211).....................: channel initialization failed MPIDI_CH3_Init(89).................: MPID_nem_init(324).................: MPID_nem_tcp_init(178).............: MPID_nem_tcp_get_business_card(425): ckpt_restart(384)..................: gethostbyname failed, zuratkul (errno 4) ************* This shows that mpich doesn't run without IPv4, which is our case... Is there a way to make it work in this situation? _______________________________________________ discuss mailing list discuss(a)mpich.org To manage subscription options or unsubscribe: https://lists.mpich.org/mailman/listinfo/discuss _______________________________________________ discuss mailing list discuss(a)mpich.org To manage subscription options or unsubscribe: https://lists.mpich.org/mailman/listinfo/discuss

1 0

Re: [mpich-discuss] Worries with ROMIO on NFS since commit b4ab2f118d
by Eric Chamberland 04 Dec '17

04 Dec '17

Hi, I have taken some time to have a relative "small" code to reproduce the problem. The good thing is that now I can extract the exact call sequence of almost any MPI I/O call we make in our code to have a pure "mpi" calls in s simple C program, independent of our in-house code. To reproduce the bug, just compile the attached file with any mpich/master since commit b4ab2f118d (nov 8) , and launch the resulting executable with 3 processes along with the second attachment (file_for_bug.data) saved in the pwd on an *NFS* path. You should see comething like this: ERROR Returned by MPI: 604040736 ERROR_string Returned by MPI: Other I/O error , error stack: ADIOI_NFS_READSTRIDED(523): Other I/O error Bad file descriptor ERROR Returned by MPI: 268496416 ERROR_string Returned by MPI: Other I/O error , error stack: ADIOI_NFS_READSTRIDED(523): Other I/O error Operation now in progress application called MPI_Abort(MPI_COMM_WORLD, 1) - process 0 application called MPI_Abort(MPI_COMM_WORLD, 1) - process 2 If you launch it on a local drive, it works. Can someone confirm it as reproduce the problem please? Moreover, if I launch it with valgrind, even on a local disk, it complains like this, on process 0: ==99023== Warning: invalid file descriptor -1 in syscall read() ==99023== at 0x53E2CB0: __read_nocancel (in /lib64/libc-2.19.so) ==99023== by 0x5041606: file_to_info_all (system_hints.c:101) ==99023== by 0x5041606: ADIOI_process_system_hints (system_hints.c:150) ==99023== by 0x50311B8: ADIO_Open (ad_open.c:123) ==99023== by 0x50161FD: PMPI_File_open (open.c:154) ==99023== by 0x400F91: main (mpich_mpiio_nfs_bug_read.c:42) Thanks, Eric On 21/11/17 02:49 PM, Eric Chamberland wrote: > Hi M. Latham, > > I have more information now. > > When I try to run my example on NFS, I have the following error code: > > error #812707360 > Other I/O error , error stack: > ADIOI_NFS_READSTRIDED(523): Other I/O error Success > > that is returned by MPI_File_read_all_begin > > When I try on a local disk, everything is fine. > > Here are all files about my actual build: > > http://www.giref.ulaval.ca/~cmpgiref/dernier_mpich/2017.11.21.05h40m02s_con… > > > http://www.giref.ulaval.ca/~cmpgiref/dernier_mpich/2017.11.21.05h40m02s_c.t… > > > http://www.giref.ulaval.ca/~cmpgiref/dernier_mpich/2017.11.21.05h40m02s_m.t… > > > http://www.giref.ulaval.ca/~cmpgiref/dernier_mpich/2017.11.21.05h40m02s_mi.… > > > http://www.giref.ulaval.ca/~cmpgiref/dernier_mpich/2017.11.21.05h40m02s_mpl… > > > http://www.giref.ulaval.ca/~cmpgiref/dernier_mpich/2017.11.21.05h40m02s_pm_… > > > http://www.giref.ulaval.ca/~cmpgiref/dernier_mpich/2017.11.21.05h40m02s_mpi… > > > Hope this help to dig further into this issue. > > Thanks, > > Eric > > > On 15/11/17 03:55 PM, Eric Chamberland wrote: >> Hi, >> >> We are compiling with mpich/master each night since august 2016... >> >> since nov 8, the mpich/master branch is buggy with our nighlty build >> tests. >> >> Here is the nov 8 config.log: >> >> http://www.giref.ulaval.ca/~cmpgiref/dernier_mpich/2017.11.08.05h36m02s_con… >> >> >> For nov 7 the configure log: >> >> http://www.giref.ulaval.ca/~cmpgiref/dernier_mpich/2017.11.07.05h36m01s_con… >> >> >> >> Since nov 8, on a specific ROMIO test, it is hanging indefinitely in >> optimized mode, and into DEBUG mode, I have a strange (yet to be >> debugged) assertion in our code. >> >> I reran the test manually, and when I wrote the results on a local >> disk, everything is fine. >> >> However, when I write over *NFS*, the test is faulty. >> >> I have not yet debugged enough through this, but, I suspect something >> related with one of: >> >> MPI_File_write_all_begin >> MPI_File_write_all_end >> MPI_File_read_all_begin >> MPI_File_read_all_end >> MPI_File_set_view >> MPI_Type_free >> >> Am I alone to see these problems? >> >> Thanks, >> Eric >> >> _______________________________________________ >> discuss mailing list discuss(a)mpich.org >> To manage subscription options or unsubscribe: >> https://lists.mpich.org/mailman/listinfo/discuss > _______________________________________________ > discuss mailing list discuss(a)mpich.org > To manage subscription options or unsubscribe: > https://lists.mpich.org/mailman/listinfo/discuss _______________________________________________ discuss mailing list discuss(a)mpich.org To manage subscription options or unsubscribe: https://lists.mpich.org/mailman/listinfo/discuss

1 0