On Wed, Apr 13, 2022 at 08:36:52AM -0700, Dave Hansen wrote: > On 4/13/22 04:30, Kirill A. Shutemov wrote: > >> 2) Fast boot; after boot, all memory will slowly but steadily get > >> accepted in the background. After a while, all memory is accepted and > >> can be signaled to user space. > ... > > Frankly, I think option 2 is the worst one. You still CPU cycles from the > > workload after boot to do the job that may or may not be needed. It is an > > half-measure that helps nobody. > > Let's not be too hyperbolic here. "Worst" is entirely subjective and it > totally depends on your perspective and what you care about. > > There are basically four options: > > * Accept everything in early boot > * Accept with deferred page free > * Accept with kthread after boot > * Accept on demand > > and four things that matter: > > * Code complexity > * Time to a shell prompt > * CPU/Memory waste > * Deterministic overhead > > Did I miss any? "Time to shell" is not equal to "time to do the job". Real workloads do stuff beyond memory allocations. But, yes, it is harder quantify. > News flash: none of the options wins on all the things that matter. > We're going to have to pick one (or maybe two). I'm also not horribly > convinced that there's a problem here worth solving, especially one that > requires surgery in the core of the buddy allocator. > > This is essentially making a performance argument: it takes too long to > boot if we go with a simpler solution. Yet, I haven't seen any data. I > think we need to go with the simplest approach(es) until there's some > actual data to guide us here. > > Here's another way to look at it: > > > https://docs.google.com/spreadsheets/d/1Fpv0Yp0CTF5_JXHR2pywvNtImTwUVGTxDMlJ5t8qiis/edit?usp=sharing The link is view-only. AFAICS, complexity of the kthread approach is on par or greater comparing to on-demand. You need coordination between allocator and the thread. It can be hard to hit right balance for the kthread between being CPU hog and not providing enough accepted memory. -- Kirill A. Shutemov