Hi Alan, On 05/13/2016 05:41 PM, One Thousand Gnomes wrote: >> My understanding is that there was a time when there was no overcommit at >> all. >> If that's the case, understanding why overcommit was introduced would be >> helpful. > > Linux always had overcommit. > > The origin of overcommit is virtual memory for the most part. In a > classic swapping system without VM the meaning of brk() and thus malloc() > is that it allocates memory (or swap). Likewise this is true of fork() > and stack extension. > > In a virtual memory system these allocate _address space_. It does not > become populated except by page faulting, copy on write and the like. It > turns out that for most use cases on a virtual memory system we get huge > amounts of page sharing or untouched space. > > Historically Linux did guess based overcommit and I added no overcommit > support way back when, along with 'anything is allowed' support for > certain HPC use cases. > > The beancounter patches combined with this made the entire setup > completely robust but the beancounters never hit upstream although years > later they became part of the basis of the cgroups. > > You can sort of set a current Linux up for definitely no overcommit using > cgroups and no overcommit settings. It works for most stuff although last > I checked most graphics drivers were terminally broken (and not just to > no overcommit but to the point you can remote DoS Linux boxes with a > suitably constructed web page and chrome browser) > > Alan >
Thanks for your comment, it certainly provides more clues and provided some history about the "overcommit" setting. I will see if we can do what we want with cgroups. Best regards, Sebastian
