chris@umcp-cs.UUCP (08/18/83)
Has anyone experimented with the vm scheduling code in 4.1BSD? Specifically, we occasionally have people run large programs that are going to take a while. Once they log out, we renice these (if the users haven't niced them already), but, alas, they still eat up most of the physical memory, thus worsening response for logged-in users. Well, limiting the "memoryuse" helps, but not much. From the code, and from experience with maxrss limited, this only affects who gets kicked out when memory is needed. It just makes the oversize process a candidate for page- or swap-out. But other stuff still tends to get kicked out: after a few seconds your login shell is no longer in core, since the system is desperate for memory. Then anytime you run something it takes forever to come back in core. So, my first idea, which I haven't attempt to implement, is to change the vm scheduler to (if there are any oversize procs) *not* kick out any *non*-oversize procs. Then only the limited processes would be affected (typically these are 2M; we only have 4M of real memory, and the kernel uses a lot of that). -- In-Real-Life: Chris Torek, Univ of MD Comp Sci UUCP: {seismo,allegra,brl-bmd}!umcp-cs!chris CSNet: chris@umcp-cs ARPA: chris.umcp-cs@UDel-Relay