On Thu, 12 Jan 2023 16:52:50 GMT, Thomas Stuefe <stu...@openjdk.org> wrote:
>> Curious, I always thought we do ArrayAllocator - using mmap for larger >> allocations - to prevent memory retention for libc variants whose allocators >> are "grabby", i.e. which don't promptly return memory to the OS on free(). >> E.g. because they only use sbrk (Solaris, AIX), or are just cautious about >> returning memory (glibc). >> >> Glibc's retention problem is only relevant for fine-grained allocations, so >> for glibc this is probably fine. This leaves at least AIX as a potential >> problem. @backwaterred, does the AIX libc malloc() still exclusively use the >> data segment ? Does free'd memory still stick to the process? >> >> (While writing this, I remember that we at SAP even rewrote Arena allocation >> to use mmap for AIX, because large compile arenas caused lasting RSS >> increase, so it has definitely been a problem in the past) > >> > To follow up on @tstuefe comment - and the one that I tried to say in the >> > bug was that we added this MmapArrayAllocate feature for some G1 marking >> > bits that used so much memory that hit the Solaris _sbrk issue. Maybe >> > @stefank and @tschatzl remember this issue. Maybe it's ok for AIX, then >> > removing this code is a good change. Maybe the G1 usages need a mmap >> > implementation though. >> >> The padding.inline.hpp usage seems to have one caller which is called once. >> The other mmap usage in G1 we can convert to mmap using a similar approach >> to zGranuleMap if that is preferred. That would then be equivalent behavior, >> it looks like the G1 code uses the page allocation granularity anyway so >> maybe keeping it mmap is the better way to go here anyway? > > My uninformed opinion (I'm not the G1 code owner) is that it would be fine to > use explicit mmap. I'd love the complexity reduction this patch brings. > @tstuefe @backwaterred I'd like to see this RFE revived. Do we know if anyone > is using the `ArrayAllocatorMallocLimit` flag in any production environment > today? > > It seems unlikely to me, as you'd need to explicitly specify > `-XX:+UnlockExperimentalVMOptions` in the command-line. > > And, if this option had been useful (for the AIX port, for example), it would > have been changed to a non-experimental (with proper `_pd` support) option > over the past 10 years. @iklam I'm fine with removing the ArrayAllocator. ------------- PR Comment: https://git.openjdk.org/jdk/pull/11931#issuecomment-1557662211
