December 2013 - discuss - mailman3.cels.anl.gov

Re: [mpich-discuss] MPI_Send/MPI_Recv - getting delayed indeffinitely
by Pavan Balaji 26 Dec '13

26 Dec '13

On Dec 25, 2013, at 8:50 PM, Madhawa Bandara <madawa911(a)gmail.com> wrote: > Thanks for the response. Actually I was using float as the data type in the program whereas MPI_DOUBLE is used in communication.. That has been the reason Ok, great. > Serializing means assigning a tag? Not exactly. Serializing is probably not the correct terminology here. The problem is what I described in my response. worker2 might send the message before worker1, in which case, it will have to wait for the master to receive worker1’s message and then process its message. If it’s not causing performance degradation for you, don’t worry about it. — Pavan -- Pavan Balaji http://www.mcs.anl.gov/~balaji

1 0

Re: [mpich-discuss] Erroneous MPI type tag warning
by Jed Brown 25 Dec '13

25 Dec '13

Markus Geimer <m.geimer(a)fz-juelich.de> writes: > Well, according to the C99 standard "Each non-bit-field member of a > structure or union object is aligned in an *implementation-defined* > manner appropriate to its type." (§6.7.2.1) It is therefore not > guaranteed that the two ints in the struct are contiguous -- which > means that, strictly speaking, int[2] is the only portable choice. > (Although I'm not aware of any compiler that would insert padding > between the two ints, i.e., in practice the struct should be fine, > too.) C11 adds the _Alignof operator and elaborates further on alignment. Although the wording is ambiguous in the section on structs, I think the intent is clear that the "implementation-defined" alignment is indeed the implementation-defined value returned by _Alignof and discussed in C11 §6.2.8 Alignment of objects: Complete object types have alignment requirements which place restrictions on the addresses at which objects of that type may be allocated. An alignment is an implementation-defined integer value representing the number of bytes between successive addresses at which a given object can be allocated. An object type imposes an alignment requirement on every object of that type: stricter alignment can be requested using the _Alignas keyword. _Alignof cannot reasonable return a value larger than sizeof, for that would break idioms such as (§6.5.3.4.7) sizeof array / sizeof array[0] Anyway, alignment rules apply everywhere, not just in structs, so there can be no difference in layout between typedef int array_2int[2]; and typedef struct {int value,loc;} struct_2int; The reiteration of "implementation-defined" in the section on structs is likely because that is a place where alignment semantics are directly observable (versus allocation of automatic/stack variables).

1 0

Fwd: Erroneous MPI type tag warning
by Junchao Zhang 25 Dec '13

25 Dec '13

On Wed, Dec 25, 2013 at 9:36 PM, Jed Brown <jedbrown(a)mcs.anl.gov> wrote: > Junchao Zhang <jczhang(a)mcs.anl.gov> writes: > > > To strictly obey MPI standard, we should declare MPI_2INT's type > attribute > > as > > > > typedef int mpich_array_mpi_2int[2]; > > static const MPI_Datatype mpich_mpi_2int > > MPICH_ATTR_TYPE_TAG_LAYOUT_COMPATIBLE (mpich_array_mpi_2int) = > > MPI_2INT; > > Why would you want this? How is this a more "strict" interpretation of > the standard? Note that for heterogeneous pairs, such as > MPI_DOUBLE_INT, the struct definition is the only one that really makes > sense and gets padding correct. > > I tried to literally interpret "MPI_TYPE_CONTIGUOUS(2, MPI_REAL, MPI_2REAL)" stated in the standard. I agree in reality it is not more strict. > > The warning happens again. In reality, struct {int value; int loc; } > should > > have the same layout as int[2]. > > Guaranteed to be the same, yes. >

1 0

Re: [mpich-discuss] Erroneous MPI type tag warning
by Jed Brown 25 Dec '13

25 Dec '13

Junchao Zhang <jczhang(a)mcs.anl.gov> writes: > To strictly obey MPI standard, we should declare MPI_2INT's type attribute > as > > typedef int mpich_array_mpi_2int[2]; > static const MPI_Datatype mpich_mpi_2int > MPICH_ATTR_TYPE_TAG_LAYOUT_COMPATIBLE (mpich_array_mpi_2int) = > MPI_2INT; Why would you want this? How is this a more "strict" interpretation of the standard? Note that for heterogeneous pairs, such as MPI_DOUBLE_INT, the struct definition is the only one that really makes sense and gets padding correct. > The warning happens again. In reality, struct {int value; int loc; } should > have the same layout as int[2]. Guaranteed to be the same, yes.

1 0

Re: [mpich-discuss] Erroneous MPI type tag warning
by Junchao Zhang 25 Dec '13

25 Dec '13

To strictly obey MPI standard, we should declare MPI_2INT's type attribute as typedef int mpich_array_mpi_2int[2]; static const MPI_Datatype mpich_mpi_2int MPICH_ATTR_TYPE_TAG_LAYOUT_COMPATIBLE (mpich_array_mpi_2int) = MPI_2INT; Then, users should write code as follows: typedef int int2[2]; int2 in, out; ... MPI_Reduce(&in, &out, 1, MPI_2INT, MPI_MAXLOC, ...); It is inconvenient. After testing the code, I found the compiler warning disappears. But if we change the code into int in[2], out[2] MPI_Reduce(in, out, 2, MPI_2INT, MPI_MAXLOC, ...); OR struct {int value; int loc; } in, out; MPI_Reduce(&in, &out, 1, MPI_2INT, MPI_MAXLOC, ...); The warning happens again. In reality, struct {int value; int loc; } should have the same layout as int[2]. In this sense, the current MPI_2INT handling in MPICH is fine. The warning reflects the compiler's weakness rather than a fault in MPICH code. --Junchao Zhang On Mon, Dec 23, 2013 at 10:55 AM, Markus Geimer <m.geimer(a)fz-juelich.de>wrote: > Hi, > > > I guess the warning has nothing to do with padding. > > Agreed. The warning is due to the type_tag attributes defined in > MPICH's 'mpi.h' header file. > > > For MPI_Reduce(in, out, 1, MPI_2INT, ...), the compiler tries to > > match int, i.e., typeof(*in), with MPI_2INT. Whatever the layout of > > MPI_2INT is defined (e.g., struct {int i1; int i2} or int[2]), the > > compiler fails at matching. If you want to use MPI_2INT, then in, out > > should be defined as of type struct {int v1; int v2}. > > Well, according to the C99 standard "Each non-bit-field member of a > structure or union object is aligned in an *implementation-defined* > manner appropriate to its type." (§6.7.2.1) It is therefore not > guaranteed that the two ints in the struct are contiguous -- which > means that, strictly speaking, int[2] is the only portable choice. > (Although I'm not aware of any compiler that would insert padding > between the two ints, i.e., in practice the struct should be fine, > too.) > > > I agree that MPI_Reduce(in, out, 1, MPI_2INT, ...) is absolutely > > right. I think the only reason for the warning is that the compiler is > > not smart enough. > > Compilers could always be smarter ;-) It would be great if you could > come up with a datatype that is -- from the compiler's perspective -- > compatible with both struct {int i1; int i2} and int[2]. Otherwise, > I suggest to remove the type check for MPI_2INT as the warning is > misleading. > > Thanks, > Markus > > > > On Sun, Dec 22, 2013 at 3:56 AM, Markus Geimer <m.geimer(a)fz-juelich.de > > <mailto:[email protected]>> wrote: > > > > Dear MPICH developers, > > > > When compiling the attached example with MPICH 3.1rc2 using Clang > 3.3, > > I get the following compiler warnings: > > > > -------------------- snip -------------------- > > mpi2int.c:17:20: warning: argument type 'int *' doesn't match > specified > > 'MPI' type tag [-Wtype-safety] > > MPI_Reduce(in, out, 1, MPI_2INT, MPI_MAXLOC, 0, MPI_COMM_WORLD); > > ^~~ ~~~~~~~~ > > mpi2int.c:17:16: warning: argument type 'int *' doesn't match > specified > > 'MPI' type tag [-Wtype-safety] > > MPI_Reduce(in, out, 1, MPI_2INT, MPI_MAXLOC, 0, MPI_COMM_WORLD); > > ^~ ~~~~~~~~ > > 2 warnings generated. > > -------------------- snip -------------------- > > > > According to the MPI standard, however, MPI_2INT is a datatype as if > > defined by > > > > MPI_Type_contiguous(2, MPI_INT, MPI_2INT) > > > > i.e., 'int[2]' should be a perfect match. This is not necessarily > true > > for the type used for comparison, > > > > struct mpich_struct_mpi_2int { int i1; int i2; }; > > > > which will only be contiguous if the compiler does not add any > padding. > > > > Is there any chance this gets fixed for the final 3.1 release? Or did > > I miss something? > > > > Thanks, > > Markus > > > > -- > Dr. Markus Geimer > Juelich Supercomputing Centre > Institute for Advanced Simulation > Forschungszentrum Juelich GmbH > 52425 Juelich, Germany > > Phone: +49-2461-61-1773 > Fax: +49-2461-61-6656 > E-mail: m.geimer(a)fz-juelich.de > WWW: http://www.fz-juelich.de/jsc/ > > > > ------------------------------------------------------------------------------------------------ > > ------------------------------------------------------------------------------------------------ > Forschungszentrum Juelich GmbH > 52425 Juelich > Sitz der Gesellschaft: Juelich > Eingetragen im Handelsregister des Amtsgerichts Dueren Nr. HR B 3498 > Vorsitzender des Aufsichtsrats: MinDir Dr. Karl Eugen Huthmacher > Geschaeftsfuehrung: Prof. Dr. Achim Bachem (Vorsitzender), > Karsten Beneke (stellv. Vorsitzender), Prof. Dr.-Ing. Harald Bolt, > Prof. Dr. Sebastian M. Schmidt > > ------------------------------------------------------------------------------------------------ > > ------------------------------------------------------------------------------------------------ > > _______________________________________________ > 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] Alternative for -nolocal
by Madhawa Bandara 25 Dec '13

25 Dec '13

Thanks a lot for the explanation Sent from my Windows phone -----Original Message----- From: "Pavan Balaji" <balaji(a)mcs.anl.gov> Sent: ‎12/‎25/‎2013 22:25 To: "discuss(a)mpich.org" <discuss(a)mpich.org> Subject: Re: [mpich-discuss] Alternative for -nolocal On Dec 24, 2013, at 10:40 AM, Madhawa Bandara <madawa911(a)gmail.com> wrote: > Is there an alternative for the -nolocal option when I execute mpiexec? -nolocal is no longer required. mpiexec takes the list of hosts you want to run the application on. If you don’t give the local host in that list, it won’t run anything locally. > What's the difference between mpirun and mpiexec? None. Both of them are symbolic links to mpiexec.hydra. — Pavan -- Pavan Balaji http://www.mcs.anl.gov/~balaji _______________________________________________ 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] MPI_Send/MPI_Recv - getting delayed indeffinitely
by Madhawa Bandara 25 Dec '13

25 Dec '13

Thanks for the response. Actually I was using float as the data type in the program whereas MPI_DOUBLE is used in communication.. That has been the reason Serializing means assigning a tag? Sent from my Windows phone -----Original Message----- From: "Pavan Balaji" <balaji(a)mcs.anl.gov> Sent: ‎12/‎25/‎2013 22:14 To: "discuss(a)mpich.org" <discuss(a)mpich.org> Subject: Re: [mpich-discuss] MPI_Send/MPI_Recv - getting delayed indeffinitely Conceptually, I don’t see anything wrong with the code snippet. There is a performance problem since you are serializing the MPI_Recv’s at the master. So it cannot receive data from worker2 before it receives data from worker1, even if worker2 is ready earlier. That’ll lose some performance, but it should not affect correctness (and for a 3 node program, you might not care about it anyway). The segmentation fault might be an issue with the serial part of your code. Typically this means you didn’t allocate enough memory somewhere. Did you try running the processes with a debugger? Instead of doing: % mpiexec -n 3 ./foo try this: % mpiexec -n 3 ddd ./foo “ddd” is a frontend for gdb and will open a different debugging screen for each process. — Pavan On Dec 23, 2013, at 8:04 PM, Madhawa Bandara <madawa911(a)gmail.com> wrote: > > Hi, > > I use mpich2 on a small cluster of 3 nodes and each node has Ubuntu 12.04 installed. I use this cluster to do the following. > > 1. Master node send some matrices to the 2 workers > 2. Workers perform some calculations and send the resulted matrices back to the master. > 3. Master perform some final calculations. > > code snippet: > > //master(taskid=0) > > MPI_Send(ha11, n / 2 * n / 2, MPI_DOUBLE, 1, 1, MPI_COMM_WORLD); //to worker 1 > MPI_Send(ha11, n / 2 * n / 2, MPI_DOUBLE, 2, 1, MPI_COMM_WORLD); //to worker 2 > > > MPI_Recv(hM1, n / 2 * n / 2, MPI_DOUBLE, 1, 2, MPI_COMM_WORLD,&status); //from worker 1 > MPI_Recv(hM2, n / 2 * n / 2, MPI_DOUBLE, 2, 2, MPI_COMM_WORLD,&status);//from worker 2 > > //final calculations using hM1,hM2 > > //worker 1 (taskid=1) > > MPI_Recv(ha11, n / 2 * n / 2, MPI_DOUBLE, 0, 1, MPI_COMM_WORLD,&status); > //does some calculations > MPI_Send(hM1, n / 2 * n / 2, MPI_DOUBLE, 0, 2, MPI_COMM_WORLD); //sends back > > > //worker 2(taskid=2) > MPI_Recv(ha11, n / 2 * n / 2, MPI_DOUBLE, 0, 1, MPI_COMM_WORLD,&status); > //does some calculations > MPI_Send(hM2, n / 2 * n / 2, MPI_DOUBLE, 0, 2, MPI_COMM_WORLD); //sends back > > > This worked fine at first, for n=128 to n=2048. But after I pushed 'n' beyond 2048 I got a segmentation fault from the worker 1. > > Since then, code works fine for the small n values. But whenever I set the value n=128 or greater, worker 1 is getting delayed indefinitely while the rest of the nodes works fine. > > What could be the reason for this? And how can I resolve this? If I have done any mistakes please point out. Thanks in advance. > -- > Regards, > H.K. Madhawa Bandara > > > > _______________________________________________ > discuss mailing list discuss(a)mpich.org > To manage subscription options or unsubscribe: > https://lists.mpich.org/mailman/listinfo/discuss -- Pavan Balaji http://www.mcs.anl.gov/~balaji _______________________________________________ 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] Alternative for -nolocal
by Pavan Balaji 25 Dec '13

25 Dec '13

On Dec 24, 2013, at 10:40 AM, Madhawa Bandara <madawa911(a)gmail.com> wrote: > Is there an alternative for the -nolocal option when I execute mpiexec? -nolocal is no longer required. mpiexec takes the list of hosts you want to run the application on. If you don’t give the local host in that list, it won’t run anything locally. > What's the difference between mpirun and mpiexec? None. Both of them are symbolic links to mpiexec.hydra. — Pavan -- Pavan Balaji http://www.mcs.anl.gov/~balaji

1 0

Re: [mpich-discuss] MPI_Send/MPI_Recv - getting delayed indeffinitely
by Pavan Balaji 25 Dec '13

25 Dec '13

Conceptually, I don’t see anything wrong with the code snippet. There is a performance problem since you are serializing the MPI_Recv’s at the master. So it cannot receive data from worker2 before it receives data from worker1, even if worker2 is ready earlier. That’ll lose some performance, but it should not affect correctness (and for a 3 node program, you might not care about it anyway). The segmentation fault might be an issue with the serial part of your code. Typically this means you didn’t allocate enough memory somewhere. Did you try running the processes with a debugger? Instead of doing: % mpiexec -n 3 ./foo try this: % mpiexec -n 3 ddd ./foo “ddd” is a frontend for gdb and will open a different debugging screen for each process. — Pavan On Dec 23, 2013, at 8:04 PM, Madhawa Bandara <madawa911(a)gmail.com> wrote: > > Hi, > > I use mpich2 on a small cluster of 3 nodes and each node has Ubuntu 12.04 installed. I use this cluster to do the following. > > 1. Master node send some matrices to the 2 workers > 2. Workers perform some calculations and send the resulted matrices back to the master. > 3. Master perform some final calculations. > > code snippet: > > //master(taskid=0) > > MPI_Send(ha11, n / 2 * n / 2, MPI_DOUBLE, 1, 1, MPI_COMM_WORLD); //to worker 1 > MPI_Send(ha11, n / 2 * n / 2, MPI_DOUBLE, 2, 1, MPI_COMM_WORLD); //to worker 2 > > > MPI_Recv(hM1, n / 2 * n / 2, MPI_DOUBLE, 1, 2, MPI_COMM_WORLD,&status); //from worker 1 > MPI_Recv(hM2, n / 2 * n / 2, MPI_DOUBLE, 2, 2, MPI_COMM_WORLD,&status);//from worker 2 > > //final calculations using hM1,hM2 > > //worker 1 (taskid=1) > > MPI_Recv(ha11, n / 2 * n / 2, MPI_DOUBLE, 0, 1, MPI_COMM_WORLD,&status); > //does some calculations > MPI_Send(hM1, n / 2 * n / 2, MPI_DOUBLE, 0, 2, MPI_COMM_WORLD); //sends back > > > //worker 2(taskid=2) > MPI_Recv(ha11, n / 2 * n / 2, MPI_DOUBLE, 0, 1, MPI_COMM_WORLD,&status); > //does some calculations > MPI_Send(hM2, n / 2 * n / 2, MPI_DOUBLE, 0, 2, MPI_COMM_WORLD); //sends back > > > This worked fine at first, for n=128 to n=2048. But after I pushed 'n' beyond 2048 I got a segmentation fault from the worker 1. > > Since then, code works fine for the small n values. But whenever I set the value n=128 or greater, worker 1 is getting delayed indefinitely while the rest of the nodes works fine. > > What could be the reason for this? And how can I resolve this? If I have done any mistakes please point out. Thanks in advance. > -- > Regards, > H.K. Madhawa Bandara > > > > _______________________________________________ > discuss mailing list discuss(a)mpich.org > To manage subscription options or unsubscribe: > https://lists.mpich.org/mailman/listinfo/discuss -- Pavan Balaji http://www.mcs.anl.gov/~balaji

1 0

Re: [mpich-discuss] MPI memory allocation.
by Pavan Balaji 25 Dec '13

25 Dec '13

I have no idea what I need to do with your benchmark since you are relying on an external tool to give performance debugging information. FWIW, there will be a memory increase in MPI_Init. That’s expected and not bad. If you post a bunch of isends, some of them get buffered locally, while others have to wait for the necessary flow-control. So it might look like your network performance has dropped. Again, this is expected. I don’t know if these are what you are seeing or if it’s a different problem. But I can’t tell from your benchmark. It might be easier for us if your application does internal profiling of memory and communication and printed it out directly. That way, I know what you are measuring more accurately. — Pavan On Dec 23, 2013, at 7:19 PM, Anatoly G <anatolyrishon(a)gmail.com> wrote: > Hi. > Do you need more info to reproduce problem? > > Anatoly. > > > On Sun, Dec 22, 2013 at 7:54 AM, Anatoly G <anatolyrishon(a)gmail.com> wrote: > Good morning Pavan. > Do you see process memory raise & network throughput low behavior? > > Anatoly. > > > On Thu, Dec 19, 2013 at 7:09 PM, Anatoly G <anatolyrishon(a)gmail.com> wrote: > No. You can see memory increase using monitor like "qps", or any other. > The stack trace that I send I get when I stop my real program (not simulation) with TotalView debugger. > if you monitor network traffic, you will see network speed which becomes low until it reaches almost 0. > > Anatoly. > > > On Thu, Dec 19, 2013 at 5:16 PM, Pavan Balaji <balaji(a)mcs.anl.gov> wrote: > > Are you printing it out? > > — Pavan > > On Dec 19, 2013, at 10:51 PM, Anatoly G <anatolyrishon(a)gmail.com> wrote: > > > Can you please remove comment from section > > /* > > // swap tag & enter blocked recv > > MPI_Status stat; > > tags[slaveIdx] = (tags[slaveIdx] == TAG1) ? TAG2 : TAG1; > > MPI_Recv(RcvBufs[slaveIdx], BUF_SZ, MPI::CHAR, slaveRank, tags[slaveIdx], MPI_COMM_WORLD, &stat); > > > > ++SlavesRcvIters[slaveIdx]; > > */ > > > > And then run it. > > Do you see memory allocation increase? > > > > Regards, > > Anatoly. > > > > > > On Thu, Dec 19, 2013 at 4:29 PM, Pavan Balaji <balaji(a)mcs.anl.gov> wrote: > > > > I’m not sure what I should look for. I ran the program and it completed fine. > > > > — Pavan > > > > On Dec 19, 2013, at 7:16 PM, Anatoly G <anatolyrishon(a)gmail.com> wrote: > > > > > Good afternoon. > > > My program enters a lot to attached stack functions. > > > Can you please explain if it's ok? > > > Did you success to execute simulation from previous mail. > > > Did you see the memory raise when MPI_Recv not in comments? > > > > > > Regards, > > > Anatoly. > > > > > > > > > On Thu, Dec 12, 2013 at 9:36 PM, Anatoly G <anatolyrishon(a)gmail.com> wrote: > > > Hi. > > > Finally, I got an additional info. > > > I build short simulation of my real application. > > > > > > The short description of real scenario. > > > I have Master + N slaves. Each slave sends to Master 2 types of messages: > > > • constant length message with predefined fields (one of it's fields is length of second message). > > > • second message - length of this message each time is different and passed in first message. > > > Master should use MPI_Irecv commands, in order to be tolerant to slaves failure (blocked MPI_Recv is blocking Master in the failure case). > > > Master executes MPI_Irecv to each slave with buffer size equal to the constant size of first message type. After receiving first type of message, Master allocates expected buffer for second message and performs receive too. This happens in endless loop for each slaves. I use MPI_Waitany to monitor all receives. > > > In order to separate messages Master & slaves use different tags (as ids) for first & second messages. > > > > > > Simulation description: > > > All passed buffers (first & second) have the same size. > > > Slave (SndSyncSlave) sends messages and swaps 2 tags between them (like 2 types of messages, but second one has constant size too). > > > Master routine (Rcv_WaitAny function) executes MPI_Irecv for first message, and after receive executes MPI_Irecv for the second one. > > > > > > In this scenario: 5 processes works fine, but if I execute 20 processes and remove comment from line "usleep(200000)" I see 800 Mbit/s on network at the test beginning, but after 1-2 second network speed become 200-300Kbit/s and not increased back any more. > > > > > > If I add MPI_Recv block in Master (remove comment from "MPI_Recv" and line around) I see that Master starts increase memory as my real application, but again on 5 processes this not happens. This is scenario used in my real application. > > > > > > Command line: mpiexec.hydra -genvall -f MpiConfigMachines.txt -launcher=ssh -n 20 mpi_rcv_any_multithread 100000 1000000 out > > > > > > where > > > 100000 - number of sends from each slave > > > 1000000 - scale to separate input from each scale (used for debug only) > > > out - prefix of output file. Each process produce out_"rank".txt file. > > > > > > MpiConfigMachines.txt - configuration file for my computers, 2 computers back to back 1 Gbit/s network. > > > > > > > > > Can you please test this case, and give me yours suggestions. > > > > > > Thank you, > > > Anatoly. > > > > > > > > > > > > On Mon, Dec 9, 2013 at 9:55 PM, Anatoly G <anatolyrishon(a)gmail.com> wrote: > > > Yes, I understand that. I"ll try to make my stand alone test closer to real application. Thank you. > > > > > > > > > On Mon, Dec 9, 2013 at 9:31 PM, Pavan Balaji <balaji(a)mcs.anl.gov> wrote: > > > > > > It sounds like MPICH is working correctly. Without a test case, it’s unfortunately quite hard for us to even know what to look for. It’s also possible that there’s a bug in your code which might be causing some bad behavior. > > > > > > — Pavan > > > > > > On Dec 9, 2013, at 1:27 PM, Anatoly G <anatolyrishon(a)gmail.com> wrote: > > > > > > > Yes, I"m actually need Fault tolerance, and it was the main reason for choosing MPICH2. I use fault tolerance for unpredictable bugs in the future. My system should survive partially. But in the regular case I just need full performance. I"m suspect that I don't use MPI correctly, but on slow rate everything works fine. The fail caused by increasing rate of MPI_Isend or increasing data buffer size. I didn't find yet any strong dependence, only main stream. > > > > > > > > Unfortunately I have a complex system which has a number of threads in each process. Part of the threads use different communicators. > > > > > > > > I try to simulate the same MPI behavior in simple stand alone test, but stand alone test works fine. It shows a full network performance, when I slow down master (on stand alone test), all slaves are stopped too and are waiting for master to continue. Can I open any MPICH log to send you the results? > > > > > > > > > > > > On Mon, Dec 9, 2013 at 8:10 PM, Pavan Balaji <balaji(a)mcs.anl.gov> wrote: > > > > > > > > Do you actually need Fault Tolerance (one of your previous emails seemed to indicate that)? > > > > > > > > It sounds like there a bug in either your application or in the MPICH stack and you are trying to trace that down, and don’t really care about fault tolerance. Is that a correct assessment? > > > > > > > > Do you have a simplified program that reproduces this error, that we can try? > > > > > > > > — Pavan > > > > > > > > On Dec 9, 2013, at 11:44 AM, Anatoly G <anatolyrishon(a)gmail.com> wrote: > > > > > > > > > No. Hardware is Ok. Master process allocates memory (check with MemoryScape doesn't show any sufficient memory allocation in my code). Then network traffic become low, and then Master process crashes w/o saving core file. I have unlimited size of core files. The same fail (w/o core) I see when I call MPI_Abort, but I don't call it. > > > > > > > > > > > > > > > On Mon, Dec 9, 2013 at 7:28 PM, Wesley Bland <wbland(a)mcs.anl.gov> wrote: > > > > > Are you actually seeing hardware failure or is your code just crashing? It's odd that one specific process would fail so often in the same way if it were a hardware problem. > > > > > > > > > > Thanks, > > > > > Wesley > > > > > > > > > > On Dec 9, 2013, at 11:15 AM, Anatoly G <anatolyrishon(a)gmail.com> wrote: > > > > > > > > > >> One more interesting fact. Each time I have a failure, the fails only master process, but slaves are still exists together with mpiexec.hydra. I thought that slaves should fail too, but slaves are live. > > > > >> > > > > >> > > > > >> On Mon, Dec 9, 2013 at 10:30 AM, Anatoly G <anatolyrishon(a)gmail.com> wrote: > > > > >> I configure MPICH-3.1rc2 build w/o "so" files. But instead of MPICH2 & MPICH-3.0.4 I get so files. What should I change in configure line to link MPI with my application statically. > > > > >> > > > > >> > > > > >> > > > > >> On Mon, Dec 9, 2013 at 9:47 AM, Pavan Balaji <balaji(a)mcs.anl.gov> wrote: > > > > >> > > > > >> Can you try mpich-3.1rc2? There were several fixes for this in this version and it’ll be good to try that out before we go digging too far into this. > > > > >> > > > > >> — Pavan > > > > >> > > > > >> On Dec 9, 2013, at 1:46 AM, Anatoly G <anatolyrishon(a)gmail.com> wrote: > > > > >> > > > > >> > With MPICH - 3.0.4 the situation repeated. It looks like MPI allocates memory for messages. > > > > >> > Can you please advice about scenario when MPI or may be TCP under MPI allocates memory due to high transfer rate? > > > > >> > > > > > >> > > > > > >> > On Mon, Dec 9, 2013 at 9:32 AM, Anatoly G <anatolyrishon(a)gmail.com> wrote: > > > > >> > Thank you very much. > > > > >> > Issend - is not so good, It can't support me Fault tolerance. If slave process fails, the master stall. > > > > >> > I tried mpich-3.0.4 with hydra-3.0.4 but my program which uses MPI Fault tolerance doesn't recognize failure of slave process, but recognizes failure with MPICH2. May be you can suggest solution? > > > > >> > I tried to use hydra from MPICH2 but link my program with MPICH3. This combination recognizes failures, but I"m not sure that such combination is stable enough. > > > > >> > Can you please advice? > > > > >> > Anatoly. > > > > >> > > > > > >> > > > > > >> > > > > > >> > On Sat, Dec 7, 2013 at 5:20 PM, Pavan Balaji <balaji(a)mcs.anl.gov> wrote: > > > > >> > > > > > >> > As much as I hate saying this — some people find it easier to think of it as “MPICH3”. > > > > >> > > > > > >> > — Pavan > > > > >> > > > > > >> > On Dec 7, 2013, at 7:37 AM, Wesley Bland <wbland(a)mcs.anl.gov> wrote: > > > > >> > > > > > >> > > MPICH is just the new version of MPICH2. We renamed it when we went past version 3.0. > > > > >> > > > > > > >> > > On Dec 7, 2013, at 3:55 AM, Anatoly G <anatolyrishon(a)gmail.com> wrote: > > > > >> > > > > > > >> > >> Ok. I"ll try both Issend, and next step to upgrade MPICH to 3.0.4. > > > > >> > >> I thought before that MPICH & MPICH2 are two different branches, when MPICH2 partially supports Fault tolerance, but MPICH not. Now I understand, that I was wrong and MPICH2 is just main version of MPICH. > > > > >> > >> > > > > >> > >> Thank you very much, > > > > >> > >> Anatoly. > > > > >> > >> > > > > >> > >> > > > > >> > >> > > > > >> > >> On Thu, Dec 5, 2013 at 11:01 PM, Rajeev Thakur <thakur(a)mcs.anl.gov> wrote: > > > > >> > >> The master is receiving too many incoming messages than it can match quickly enough with Irecvs. Try using MPI_Issend instead of MPI_Isend. > > > > >> > >> > > > > >> > >> Rajeev > > > > >> > >> > > > > >> > >> On Dec 5, 2013, at 2:58 AM, Anatoly G <anatolyrishon(a)gmail.com> wrote: > > > > >> > >> > > > > >> > >> > Hello. > > > > >> > >> > I"m using MPICH2 1.5. > > > > >> > >> > My system contains master and 16 slaves. > > > > >> > >> > System uses number of communicators. > > > > >> > >> > The single communicator used for below scenario: > > > > >> > >> > Each slave sends non-stop 2Kbyte data buffer using MPI_Isend and waits using MPI_Wait. > > > > >> > >> > Master starts with MPI_Irecv to each slave > > > > >> > >> > Then in endless loop: > > > > >> > >> > MPI_Waitany and MPI_Irecv on rank returned by MPI_Waitany. > > > > >> > >> > > > > > >> > >> > Another communicator used for broadcast communication (commands between master + slaves), > > > > >> > >> > but it's not used in parallel with previous communicator, > > > > >> > >> > only before or after data transfer. > > > > >> > >> > > > > > >> > >> > The system executed on two computers linked by 1Gbit/s Ethernet. > > > > >> > >> > Master executed on first computer, all slaves on other one. > > > > >> > >> > Network traffic is ~800Mbit/s. > > > > >> > >> > > > > > >> > >> > After 1-2 minutes of execution, master process starts to increase it's memory allocation and network traffic becomes low. > > > > >> > >> > This memory allocation & network traffic slow down continues until fail of MPI, > > > > >> > >> > without core file save. > > > > >> > >> > My program doesn't allocate memory. Can you please explain this behaviour. > > > > >> > >> > How can I cause MPI to stop sending slaves if Master can't serve such traffic, instead of memory allocation and fail? > > > > >> > >> > > > > > >> > >> > > > > > >> > >> > Thank you, > > > > >> > >> > Anatoly. > > > > >> > >> > > > > > >> > >> > P.S. > > > > >> > >> > On my stand alone test, I simulate similar behaviour, but with single thread on each process (master & hosts). > > > > >> > >> > When I start stand alone test, master stops slaves until it completes accumulated data processing and MPI doesn't increase memory allocation. > > > > >> > >> > When Master is free slaves continue to send data. > > > > >> > >> > _______________________________________________ > > > > >> > >> > 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 > > > > >> > > > > > >> > -- > > > > >> > Pavan Balaji > > > > >> > http://www.mcs.anl.gov/~balaji > > > > >> > > > > > >> > _______________________________________________ > > > > >> > 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 > > > > >> > > > > >> -- > > > > >> Pavan Balaji > > > > >> http://www.mcs.anl.gov/~balaji > > > > >> > > > > >> _______________________________________________ > > > > >> 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 > > > > > > > > -- > > > > Pavan Balaji > > > > http://www.mcs.anl.gov/~balaji > > > > > > > > _______________________________________________ > > > > 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 > > > > > > -- > > > Pavan Balaji > > > http://www.mcs.anl.gov/~balaji > > > > > > _______________________________________________ > > > discuss mailing list discuss(a)mpich.org > > > To manage subscription options or unsubscribe: > > > https://lists.mpich.org/mailman/listinfo/discuss > > > > > > > > > > > > <backtrace2.txt>_______________________________________________ > > > discuss mailing list discuss(a)mpich.org > > > To manage subscription options or unsubscribe: > > > https://lists.mpich.org/mailman/listinfo/discuss > > > > -- > > Pavan Balaji > > http://www.mcs.anl.gov/~balaji > > > > _______________________________________________ > > 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 > > -- > Pavan Balaji > http://www.mcs.anl.gov/~balaji > > _______________________________________________ > 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 -- Pavan Balaji http://www.mcs.anl.gov/~balaji

1 0