!- Converted with LaTeX2HTML 0.6.4 (Tues Aug 30 1994) by Nikos Drakos (email@example.com), CBLU, University of Leeds ->
How many tasks each pvmd can manage is limited by two factors: the number of processes allowed a user by the operating system, and the number of file descriptors available to the pvmd. The limit on processes is generally not an issue, since it doesn't make sense to have a huge number of tasks running on a uniprocessor machine.
Each task consumes one file descriptor in the pvmd, for the pvmd-task TCP stream. Each spawned task (not ones connected anonymously) consumes an extra descriptor, since its output is read through a pipe by the pvmd (closing stdout and stderr in the task would reclaim this slot). A few more file descriptors are always in use by the pvmd for the local and network sockets and error log file. For example, with a limit of 64 open files, a user should be able to have up to 30 tasks running per host.
The pvmd may become a bottleneck if all these tasks try to talk to one another through it.
The pvmd uses dynamically allocated memory to store message packets en route between tasks. Until the receiving task accepts the packets, they accumulate in the pvmd in an FIFO procedure. No flow control is imposed by the pvmd: it will happily store all the packets given to it, until it can't get any more memory. If an application is designed so that tasks can keep sending even when the receiving end is off doing something else and not receiving, the system will eventually run out of memory .