What is gang scheduling?
Gang Scheduling is described a variety of ways. Generally the consistent thread is that a GS gives a process all the processors at the same time (or none for a time slice). This is most helpful for “scientific apps” because the most common set up is something like do i=1, bignum stuff more stuff lots more stuff end do the obvious decomposition is bignum/nproc statically allocated. Stuff and friends take very close to the same time per chunk, so if you get lucky it all happens in one chime (viz. one big clock). Else it takes precisely N chimes with no leftovers. When unlucky, it’s N chimes + cleanup for stragglers. Virtually all supercomputers do this, they may not even bother to give it a special name. SGI makes this explicit (and supported). On SPARC/Solaris there is no way for the compiler to know if we’ll get the processors requested or when. So you can suffer multiple chime losses quite easily. One can reallocate processor/code on the fly, but with increased overhead. =============
