> It never seemed to. > > But of course if you have procs with same pid, the collisions are obvious. > > So, do the easy thing: > > for all nodes, mount them at > /proc/localhost > /proc/hostname/whatever > > Then modify ps (takes about 5 minutes) so it iterates over /proc/* > where * is a set of host names. > > now you can do fun stuff > slay node8/mpirun | rc > slay node*/mpirun | rc > > There's a lot of good stuff in there if you want to use it ... I > actually implemented all this a few years back when Vic did hist first > xcpu code. It was really nice.
the trivial solution on your hardware would be to partition the pid space, wouldn't it. just have 64bit pids? let each machine start at a 1<<32 boundary? four billion machines ought with four billion processes each ought to be enough for anyone. - erik
