User Program Restrictions under SCore
The following restrictions apply to user programs run under SCore:
- SCore cluster system software is not sucure enough to be
exposed to the internet. Place cluster(s) and related servers
inside of a firewall.
- There is no way to specify an exact number of nodes for which
user program to be allocated. Thus, it is not a good idea for
a user program to assume a certain number of nodes to be
allocated.
- SIGQUIT is reserved for checkpointing. SIGCONT is reserved for
gang scheduling.
- SCore users must have the same user ID and group IDs on all
cluster hosts and on hosts where
scrun is
invoked.
- All environment variables at the time of the invocation of the
scrun front-end process are copied to each
cluster host on which the user parallel program will run.
- Do not attach a debugger to a user process running on a cluster host.
Use the
debug option and
sc_inspectme() SCore-D system call instead.
- User program binary files must have read permission so that
SCore-D can distribute the binary file to the compute hosts
before invoking the user program on the cluster. Thus, users' binary
files can exist on a local file system of a host where
scrun is invoked.
- AFS, used as a network file system for a cluster wide file system,
is not supported. This is because an AFS server may go down when a
cluster consists of more than one hundred hosts. However, user
program binary files can be located on an AFS file system.
- If the current directory of a user program is on an NFS file system,
then core files may not be properly produced when two or more
processes raise exceptions simultaneously. This is why automatic
debugger attachment (the
debug option) for exceptions is
supported in MPC++ and MPICH-SCore.
