> From: Bruce Richardson [mailto:bruce.richard...@intel.com] > Sent: Monday, 8 April 2024 10.11 > > On Sun, Apr 07, 2024 at 12:04:24AM +0200, Morten Brørup wrote: > > > From: Patrick Robb [mailto:pr...@iol.unh.edu] > > > Sent: Saturday, 6 April 2024 21.21 > > > > > > On Sat, Apr 6, 2024 at 4:47 AM Morten Brørup <m...@smartsharesystems.com> > > > wrote: > > > > > > > > > > > > This change seems very CPU specific. > > > > > > > > E.g. in x86 32-bit mode, the hugepage size is 4 MB, not 2 MB. > > > > > > > > I don't know the recommended hugepage size on other architectures. > > > > > > > > > > Thanks Morten, that's an important insight which we weren't aware of > > > when we initially discussed this ticket. > > > > > > We read on some dpdk docs that 1gb hugepages should be set at boot (I > > > think the reason is because that's when you can guarantee there is gbs > > > of contiguous available memory), and that after boot, only 2gb > > > hugepages should be set. I assume that even for other arches which > > > don't support the 2mb pages specifically, we still want to allocate > > > hugepages using the smallest page size possible per arch (for the same > > > reason). > > > > Correct; for very large hugepages, they need to be set at boot. > > I don't remember why, but that's the way the Linux kernel works. > > 2 MB is way below that threshold, and 1 GB is above. > > > > You can also not set nr_overcommit_hugepages for those very large hugepages, > only nr_hugepages. > > > > > > > > So I think we can add some dict which stores the smallest valid > > > hugepage size per arch. Then during config init, use the config's arch > > > value to determine that size, and set the total hugepages allocation > > > based on that size and the hugepages count set in the conf.yaml. Or > > > maybe we can store the list of all valid hugepgage sizes per arch > > > (which are also valid to be set post boot), allow for a new > > > hugepage_size value on the conf.yaml, validate the input at config > > > init, and then set according to those values. I prefer the former > > > option though as I don't think the added flexibility offered by the > > > latter seems important. > > > > I agree; the former option suffices. > > > > A tiny detail... > > ARM supports multiple (4, I think) different hugepage sizes, where the > smallest size is 64 KB. > > So you might want to choose another hugepage size than the smallest; but I > still agree with your proposed concept of using one specific hugepage size per > arch. > > > > Like x86_64, ARM also supports 2 MB and 1 GB hugepage sizes, and 2 MB > hugepages is also the typical default on Linux. > > > > I don't know which hugepage sizes are supported by other architectures. > > It might only be 32-bit x86 that needs a different hugepage size than 2 MB. > > > > Even for 32-bit x86, I think most distros now use PAE mode to allow physical > addresses >4GB, so even then 32-bit hugepages are 2MB rather than 4MB.
In order to save you from more work on this patch, we could assume/hope that all architectures support 2 MB hugepages, and drop the discussed per-architecture size dict. At least for until the assumption doesn't hold true anymore. Then I only have two further comments to the patch: 1. Don't increase from 256 to 2560 hugepages (in dts/conf.yaml), unless there is a reason for it. If there is, please mention the increase and reason in the patch description. 2. Don't remove the size parameter from _configure_huge_pages() (in /dts/framework/testbed_model/linux_session.py); it might be useful later.
