[mpich-discuss] Error while connecting to host

Takahiro Someji Takahiro_Someji at hitachi-metals.co.jp
Sat Mar 13 02:35:19 CST 2010


Hi.
  Thank you very much for your work.
  I also checked again. Then, it operated normally as you told. I do not 
understand the cause.
  Because the difference of the setting of PC might be a cause, I have 
been tested until now using two sets of PCs (setting is different).
  Moreover, I am confirming the same phenomenon at rev1.0.7. Are the 
domain setting of PC (workgroup or ActiveDirectory) and the setting such 
as LAN Manager related?
  I want to investigate the cause of not operating normally.

Regards,
Someji

(2010/03/13 0:35), jayesh at mcs.anl.gov wrote:
> Hi,
>   With 1.3a1 (and 1.2.1p1) you should also be able to start MPI processes as described in the "Debugging jobs by starting them manually" section of the developer's guide. I verified that the setup works (on 1.2.1p1&  should work on 1.3 series too) on my local machine. Please follow the steps below to start MPI processes manually on the local machine,
>
> 1) Open two command prompts, command_prompt_1&  command_prompt_2
> 2) On command_prompt_1 please set the PMI environment as follows,
>
>     set PMI_ROOT_HOST=000520p80812.ad.hitachi-metals.co.jp
>     set PMI_ROOT_PORT=9222
>     set PMI_RANK=0
>     set PMI_SIZE=2
>     set PMI_KVS=mpich2
>     set PMI_ROOT_LOCAL=1
>
> 3) On command_prompt_2 please set the PMI environment as follows,
>
>     set PMI_ROOT_HOST=000520p80812.ad.hitachi-metals.co.jp
>     set PMI_ROOT_PORT=9222
>     set PMI_RANK=1
>     set PMI_SIZE=2
>     set PMI_KVS=mpich2
>     set PMI_ROOT_LOCAL=1
>
> 4) Run your MPI process (by typing the executable name at the command prompt - don't use mpiexec to launch your program) on the command_prompt_1 . (Note: The process with PMI_RANK=0 should be run before other ranks)
>
> 5) Run your MPI process on command_prompt_2
>
>   Let us know if it works.
>
> (PS: The environment variable PMI_RANK determines the rank of the current MPI process and the var PMI_SIZE determines the total number of MPI processes in the job)
> Regards,
> Jayesh
>
>
> ----- Original Message -----
> From: "Takahiro Someji"<Takahiro_Someji at hitachi-metals.co.jp>
> To: "Jayesh Krishna"<jayesh at mcs.anl.gov>
> Cc: mpich-discuss at mcs.anl.gov
> Sent: Friday, March 12, 2010 2:42:39 AM GMT -06:00 US/Canada Central
> Subject: Re: [mpich-discuss] Error while connecting to host
>
> Hi.
>    When rev1.3a1 was installed, it worked well. (At rev1.0.3, when
> setting up as the manual, it operated normally.)
>
>    However, there is a problem by this method.
>    It is a problem that a port number and a domain name, kvs name change
> whenever it starts a PMI server.
>    It is troublesome to start a PMI server, and to set up these
> parameters again at every time.
>    Are there better methods.
>
>    
>>    Let us know if it works. Meanwhile, why do you want to start your MPI processes manually ?
>>      
>    I am developing image-processing system using MPI.
>    In order to obtain high performance, I adopted the distributed
> processing by multi PCs and MPI.
>    The program is developed by many men. And the contents of the program
> differ for every PC (different progarams are operated on each PCs).
>    Therefore, everybody debug individually at the time of debugging.
> Finally, it is necessary to build as one application.
>    At the time of execution, it starts at once from a host. However, I
> want to start separately at the time of debugging and to perform MPI
> communication.
>    If there is a better idea, please let me know.
>
> Regards,
> Someji
>
> (2010/03/12 13:48), Jayesh Krishna wrote:
>    
>> Hi,
>>    You are very close. When specifying the environment for MPI processes using SMPD as the process manager you should not specify the PMI_LOCAL environment variable.
>>    Try the following steps to start MPI processes manually,
>>
>> 1) Run "mpiexec -pmiserver 2" on the command prompt. The command would output the host, port and kvs information in the order.
>>
>>    >D:\temp>mpiexec -pmiserver 2
>>    >000520P80812.ad.hitachi-metals.co.jp (set PMI_HOST&   PMI_ROOT_HOST to this hostname)
>>    >2612 (set PMI_PORT&   PMI_ROOT_PORT to this port number)
>>    >65B8416A-4669-436b-A342-9FD0FF3357F9 (set PMI_KVS&   PMI_DOMAIN to this hostname)
>>
>> 2) Now open two command prompts - command_prompt_1&   command_prompt_2
>>
>> 3) On command_prompt_1 set the PMI environment as follows,
>>    >set PMI_ROOT_HOST=000520p80812.ad.hitachi-metals.co.jp
>>    >set PMI_HOST=000520p80812.ad.hitachi-metals.co.jp
>>    >set PMI_ROOT_PORT=2612
>>    >set PMI_PORT=2612
>>    >set PMI_RANK=0
>>    >set PMI_SIZE=2
>>    >set PMI_KVS=65B8416A-4669-436b-A342-9FD0FF3357F9
>>    >set PMI_DOMAIN=65B8416A-4669-436b-A342-9FD0FF3357F9
>>
>> 4) On command_prompt_2 set the PMI environment as follows,
>>    >set PMI_ROOT_HOST=000520p80812.ad.hitachi-metals.co.jp
>>    >set PMI_HOST=000520p80812.ad.hitachi-metals.co.jp
>>    >set PMI_ROOT_PORT=2612
>>    >set PMI_PORT=2612
>>    >set PMI_RANK=1
>>    >set PMI_SIZE=2
>>    >set PMI_KVS=65B8416A-4669-436b-A342-9FD0FF3357F9
>>    >set PMI_DOMAIN=65B8416A-4669-436b-A342-9FD0FF3357F9
>>
>> 5) Now run the MPI program (by typing the name of the executable at the command prompt) on command_prompt_1 and command_prompt_2
>>
>>    If the above steps don't work please try the preview release of MPICH2 (1.3a1 - there has been several bug fixes added after the 1.2.1p1 release) and see if it works.
>>    Let us know if it works. Meanwhile, why do you want to start your MPI processes manually ?
>>
>> Regards,
>> Jayesh
>>
>> ----- Original Message -----
>> From: "Takahiro Someji"<Takahiro_Someji at hitachi-metals.co.jp>
>> To: jayesh at mcs.anl.gov
>> Cc: mpich-discuss at mcs.anl.gov
>> Sent: Thursday, March 11, 2010 10:30:39 PM GMT -06:00 US/Canada Central
>> Subject: Re: [mpich-discuss] Error while connecting to host
>>
>> Hello.
>>     Sorry, I was not able to understand your explanation completely. Since
>> I do not know well about the PMI, please teach me more closely.
>>     If there is another method of launch processes manually without using
>> PMI, please let me know.
>>
>>     Well, I tried.
>>
>> I operated as shown below in one command window.
>>    >D:\temp>mpiexec -pmiserver 2
>>    >000520P80812.ad.hitachi-metals.co.jp
>>    >2612
>>    >65B8416A-4669-436b-A342-9FD0FF3357F9
>>
>> Next, in anther two window (Host&   Sub Window), I set up PMI
>> environment, as shown below.
>>    >set PMI_ROOT_HOST=000520p80812.ad.hitachi-metals.co.jp
>>    >set PMI_HOST=000520p80812.ad.hitachi-metals.co.jp
>>    >set PMI_ROOT_PORT=2612
>>    >set PMI_PORT=2612
>>    >set PMI_LOCAL=1
>>    >set PMI_RANK=%1
>>    >set PMI_SIZE=2
>>    >set PMI_KVS=65B8416A-4669-436b-A342-9FD0FF3357F9
>>    >set PMI_DOMAIN=65B8416A-4669-436b-A342-9FD0FF3357F9
>>
>> Next , I started the program with cmd "program.exe" in Host window&   Sub
>> window.
>> Then, the program was terminated for the fatal error (access violation).
>>      (Is this a fundamental problem?)
>>
>> Next, I set PMI_ROOT_PORT and PMI_PORT to 9222.
>> Then, the result is "ERROR:Error while connecting to host......".
>>
>> What is bad?.
>>
>> Regars,
>> Someji
>>
>>
>> (2010/03/12 1:05), jayesh at mcs.anl.gov wrote:
>>
>>      
>>> Hi,
>>>     Try also setting the PMI_DOMAIN value to the PMI_KVS value.
>>>     If you just want to start your programs manually (launch processes manually) and still use SMPD as a PMI server I would recommend using the "-pmiserver" option of mpiexec (mpiexec -pmiserver 2 =>    Starts an instance of SMPD - the port/kvs/domain values are printed out - and you can launch your MPI processes with the provided PMI environment. I set PMI_SIZE, PMI_RANK, PMI_KVS, PMI_DOMAIN, PMI_PORT, PMI_ROOT_PORT, PMI_HOST, PMI_ROOT_HOST env).
>>>     Let us know if it works for you.
>>> Regards,
>>> Jayesh
>>>
>>> (PS: When connecting to the process manager the default values of the kvs won't work.)
>>> ----- Original Message -----
>>> From: "Takahiro Someji"<Takahiro_Someji at hitachi-metals.co.jp>
>>> To:jayesh at mcs.anl.gov,mpich-discuss at mcs.anl.gov
>>> Sent: Wednesday, March 10, 2010 6:53:22 PM GMT -06:00 US/Canada Central
>>> Subject: Re: [mpich-discuss] Error while connecting to host
>>>
>>> Hello.
>>>
>>>      I tried as your proposal. However, the result did not change.
>>>
>>>      I tried a setup of PMI_ROOT_PORT=8676 as smpd default port number in
>>> Host window&    Sub window.
>>>      Then, the program was terminated for the fatal error.
>>>      Is communication with smpd blocked? Or is this a unique phenomenon by
>>> japanese OS?
>>>
>>> Regards,
>>> Someji
>>>
>>> (2010/03/11 0:08),jayesh at mcs.anl.gov  wrote:
>>>
>>>
>>>        
>>>> Hi,
>>>>      Try adding PMI_HOST (with value of PMI_ROOT_HOST) and PMI_PORT (with value of PMI_ROOT_PORT) into your PMI environment and see if it works. Make sure that you start the process with rank=0 before other processes.
>>>>      Are you trying to debug your program (Another way to debug your code would be to attach to the MPI process using a debugger)?
>>>>
>>>> Regards,
>>>> Jayesh
>>>> ----- Original Message -----
>>>> From: "染次 孝博"<Takahiro_Someji at hitachi-metals.co.jp>
>>>> To:mpich-discuss at mcs.anl.gov
>>>> Sent: Tuesday, March 9, 2010 11:51:40 PM GMT -06:00 US/Canada Central
>>>> Subject: [mpich-discuss] Error while connecting to host
>>>>
>>>>
>>>> Hi.
>>>>
>>>> I am developing the manual start program. (WindowsXP sp3, MPICH2-1.2.1p1, VisualStudio2008 sp1, C++)
>>>> Program is very simple as below.
>>>>
>>>> ************************
>>>> *int main(int argc, char* argv[])
>>>> *{
>>>> * int numprocs,myid,namelen,i;
>>>> * char processor_name[256];
>>>> *
>>>> * MPI_Init(&argc,&argv);
>>>> * MPI_Comm_size(MPI_COMM_WORLD,&numprocs);
>>>> * MPI_Comm_rank(MPI_COMM_WORLD,&myid);
>>>> * MPI_Get_processor_name(processor_name,&namelen);
>>>> *
>>>> * for (i=0 ; i<argc ; i++)
>>>> * {
>>>> * printf("arg%d:%s\n",i,argv[i]);
>>>> * }
>>>> * printf("numprocs:%d\n",numprocs);
>>>> * printf("myid:%d\n",myid);
>>>> * printf("processor name:%s\n",processor_name);
>>>> *
>>>> * MPI_Finalize();
>>>> * return 0;
>>>> *}
>>>> ************************
>>>>
>>>> The result of the program with "mpiexec -n 2 program.exe" is below.
>>>>
>>>> *************************
>>>> *arg0:D:\temp\MPITest1\Debug\mpitest1.exe
>>>> *numprocs:2
>>>> *myid:1
>>>> *processor name:HOSTPC
>>>> *arg0:D:\temp\MPITest1\Debug\mpitest1.exe
>>>> *numprocs:2
>>>> *myid:0
>>>> *processor name:HOSTPC
>>>> *************************
>>>>
>>>> Next, I tried to start of the two programs manually with one PC.
>>>> As shown in the manual "Debugging jobs by starting them manually" , it set up as follows.
>>>> In Host window:
>>>> set PMI_ROOT_HOST=HOSTPC
>>>> set PMI_RANK=0
>>>> set PMI_ROOT_PORT=9222
>>>> set PMI_ROOT_LOCAL=1
>>>> set PMI_SIZE=2
>>>> set PMI_KVS=mpich2
>>>> In Sub window:
>>>> set PMI_ROOT_HOST=HOSTPC
>>>> set PMI_RANK=1
>>>> :
>>>> : same as Host
>>>>
>>>> Then , I started this program with cmd "program.exe" in Host window&     Sub window.
>>>> As a result, the following errors occurred and it did not operate well.
>>>>
>>>> [01:3156]..ERROR:Error while connecting to host,
>>>> Can not connected because refused target computer (10061)
>>>> [01:3156]..ERROR:Connect on sock (host=000520p80812.ad.hitachi-metals.co.jp, por
>>>> t=9222) failed, exhaused all end points
>>>> SMPDU_Sock_post_connect failed.
>>>> [0] PMI_ConnectToHost failed: unable to post a connect to 000520p80812.ad.hitach
>>>> i-metals.co.jp:9222, error: Undefined dynamic error code
>>>> uPMI_ConnectToHost returning PMI_FAIL
>>>> [0] PMI_Init failed.
>>>> Fatal error in MPI_Init: Other MPI error, error stack:
>>>> MPIR_Init_thread(394): Initialization failed
>>>> MPID_Init(103).......: channel initialization failed
>>>> MPID_Init(374).......: PMI_Init returned -1
>>>>
>>>>
>>>> Please let me know solution.
>>>>
>>>> -- Thank you.
>>>>
>>>> -----------------------------------------------
>>>> Takahiro Someji , Senior Engineer
>>>>
>>>> Hitachi Metals Ltd. Production System Lab.
>>>> 6010, Mikajiri
>>>> Kumagaya city,Saitama pref. JAPAN
>>>> zip: 360-0843
>>>>
>>>> phone: +81-485-31-1720
>>>> fax: +81-485-33-3398
>>>> eMail:takahiro_someji at hitachi-metals.co.jp
>>>> web:http://www.hitachi-metals.co.jp
>>>> -----------------------------------------------
>>>>
>>>>          
>>
>>      
>
>    




More information about the mpich-discuss mailing list