| PVM_SHMD(1PVM) | PVM Version 3.4 | PVM_SHMD(1PVM) |
pvm_shmd - PVM shared memory (2) daemon
pvm_shmd [ -options ] [ hostfile ]
Pvm_shmd is a daemon process which maintains and handles the usage of shared memory resources (memory segments, semaphores, message queues) on a particular PVM host. One pvm_shmd must run on each host in a virtual machine that wants to use the shmd shared memory message passing layer. The shmd message passing layer allows local processes to a host to use shared memory for message passing on that host ONLY. See the shmd directory README for further notes and restrictions.
The daemon can be started manually as with any other PVM task. Only one pvm_shmd can be active per host. If the pvm_shmd detects another pvm_shmd it will shutdown.
The pvm_shmd can be killed from the PVM console by using the reset command. Local and remote pvm_shmds can also be started from the PVM console program pvm.
The name of the daemon executable is pvm_shmd.
The pvm_shmd will attempt to create upto MAXSEG shared memory segments of upto MAXPAGES memory pages in size. The size of each memory page is architecture dependent(see getpagesize(3C)). The segments are then numbered 0..N-1 where N is the number of segments created. The first segement (0) has the shared memory control structure placed in its head. This structure allows alien processes to located the other shared memory segments and any required controlling information.
The message passing layer, allocates memory from these segments asynchronously without any interaction with the pvm_shmd using semaphores to protect data during updates to any associated structures. For each segment there is an associated page map of which processes have currently locked a page. Each page map for a segment has a separate semaphore protecting it. The semaphores are accessed with the SEM_UNDO flag set so that if a process holding a semaphore should die, the OS (should) reset the semaphore automatically, thus allowing any waiting/blocked processes to continue.
The pvm_shmd only maintains the segments and their allocation page map(s). Thus if a process allocated pages in a segment and then exits, it is the pvm_shmd that detects this and then frees the allocated pages.
The pvm_shmd can have its status checked at any time by using the pvm_shmd_stat process to kick it into reporting onto either stdio or the pvm log file (pvml.uid) its internal state.
The shared memory and all associated processes can be cleared by sending the pvm_shmd a HUP signal.
Sending the pvm_shmd a TERM signal will just cause it to clear any shared resources and then exit.
The pvm_shmd cannot catch the KILL signal.
If the pvm_shmd is killed without clearing all of its shared resources these can be cleared by calling ipcfree which resides in the pvm3/lib directory.
The pvm_shmd handles resources that are used by special versions of pvm_psend() and pvm_precv() stored in the libpvmshmd.a library. Thus to use these facilities, applications have to link to this library instead of the usual libpvm3.a library.
The following options may be specified on the command line when starting the pvm_shmd:
Remember that pvm_shmd allocated memory from the VM available on the machine. Allocating more segements improves performance as there is less sharing of segments (semaphores for their page maps). Although you must remember to leave some memory available for normal program and OS system usage, as the pvm_shmd allocated memory is *ONLY* used for message passing.
$PVM_ROOT/lib/$PVM_ARCH/pvm_shmd PVM shared memory daemon executable
$PVM_ROOT/shmd/shmd.h Shared Memory hard limits header file
/tmp/pvml.uid Pvmd runtime error log
pvm(1PVM), pvmd(1PVM), getpagesize(3C), ipcs(1), msgctl(2), semctl(2), shmctl(2), signal(5)
| 11 March, 1999 | G.E.Fagg |