Справочник Пользователя для Escali 4.4

Section:  

mpid opens a socket and assigns a predefined mpid port number (see /etc/services for more 
information), to the end point. If mpid is terminated abnormally, the mpid port number cannot 
be re-used until a system defined timer has expired. To resolve:
‹ Use netstat -a | grep mpid to observe when the socket is released. When the socket is 

released, restart mpid again.

B-1.2.1 Bad clean up
V

‹ Check for mpi-processes on the nodes using /opt/scali/bin/scaps.
‹ Use /opt/scali/sbin/scidle 
‹ Use /opt/scali/bin/scash to check for leftover shared memory segments on all nodes 

(ipcs for Solaris and Linux).

Note: core dumping takes time.

B-1.2.2 Space overflow
V

‹ The mpimon pool-size specifications are too large, and must be reduced.

B-1.3.1 Core dump
V

‹ Use a debugger to locate the point of violation. The application may need to be recompiled 

to include symbolic debug information (-g for most compilers).

‹ Define SCAMPI_INSTALL_SIGSEGV_HANDLER=1 and attach to the failing process with the 

debugger.

V

‹ Check if every send has a matching receive.

V

‹ If the application has a large degree of asynchronicity, try to increase the channel-size. 

Further information is available in “Communication buffer adaption: If the communication 
behaviour of the application is known, explicitly providing buffersize settings to mpimon, to 
match the requirement of the application, will in most cases improve performance.  
Example: Application sending only 900 bytes messages. Set channel_inline_threshold 964 
(64 added for alignment) and increase the channel-size significantly (32-128 k). Setting 
eager_size 1k and eager_count high (16 or more). If all messages can be buffered, the 
transporter-{size, count} can be set to low values to reduce shared memory consumption.” 
on page 47. 

V

‹ Investigate the core file, or rerun the program in a debugger.