Hi! When trying to adapt GridScheduler's (former SGE) Java DRMAA wrappers to Slurms C DRMAA plugin, I have discovered something peculiar about the way Slurm's plugins (for authenticaction, etc.) are loaded:
When using Slurm's libdrmaa.so and its example C program, everything is working correctly. But when trying to run another test program from Java with a JNI wrapper library to libdrmaa, I get "undefined symbol" errors in auth_munge.so for symbols defined in libslurm.so. After some digging, I think I understand why this is: I use Munge for authentication and when libslurm.so is used it tries to load auth_munge.so. It is loaded at run-time (with ldopen(...) ) and in turn uses symbols from libslurm.so, but without linking to it. The way this seems to function in the regular pure-C way is that libslurm is loaded in "global mode" (the RTLD_GLOBAL flag to ldopen, or equivalent), making all symbols in it available implicilty. Hence it works. But from JNI, it is loaded in "local mode" (RTLD_LOCAL) and hence auth_munge cannot find the symbols from libslurm. The same issue seems to exist in most plugins. Is there some reason for why the plugins are not linked to libslurm when built, or is my build not configured correctly? Else, it would seem that library files should link to all other libraries they use. (There seems to be a way to force libslurm to be loaded in global mode, by explicitly loading libdrmaa (which in turn loads libslurm) in the JNI on-load hook, but I don't now id this has any consequences I haven't discovered yet...) Regards, /Sebastian Gröhn --- Sebastian Gröhn, Research Assistant Dept. of Computing Science, Umeå University
