Re: [PATCH v10 00/11] Free some vmemmap pages of HugeTLB page

From: David Hildenbrand
Date: Thu Dec 17 2020 - 07:19:44 EST


On 17.12.20 13:12, Muchun Song wrote:
> Hi all,
>
> This patch series will free some vmemmap pages(struct page structures)
> associated with each hugetlbpage when preallocated to save memory.
>
> In order to reduce the difficulty of the first version of code review.
> From this version, we disable PMD/huge page mapping of vmemmap if this
> feature was enabled. This accutualy eliminate a bunch of the complex code
> doing page table manipulation. When this patch series is solid, we cam add
> the code of vmemmap page table manipulation in the future.
>
> The struct page structures (page structs) are used to describe a physical
> page frame. By default, there is a one-to-one mapping from a page frame to
> it's corresponding page struct.
>
> The HugeTLB pages consist of multiple base page size pages and is supported
> by many architectures. See hugetlbpage.rst in the Documentation directory
> for more details. On the x86 architecture, HugeTLB pages of size 2MB and 1GB
> are currently supported. Since the base page size on x86 is 4KB, a 2MB
> HugeTLB page consists of 512 base pages and a 1GB HugeTLB page consists of
> 4096 base pages. For each base page, there is a corresponding page struct.
>
> Within the HugeTLB subsystem, only the first 4 page structs are used to
> contain unique information about a HugeTLB page. HUGETLB_CGROUP_MIN_ORDER
> provides this upper limit. The only 'useful' information in the remaining
> page structs is the compound_head field, and this field is the same for all
> tail pages.
>
> By removing redundant page structs for HugeTLB pages, memory can returned to
> the buddy allocator for other uses.
>
> When the system boot up, every 2M HugeTLB has 512 struct page structs which
> size is 8 pages(sizeof(struct page) * 512 / PAGE_SIZE).
>
> HugeTLB struct pages(8 pages) page frame(8 pages)
> +-----------+ ---virt_to_page---> +-----------+ mapping to +-----------+
> | | | 0 | -------------> | 0 |
> | | +-----------+ +-----------+
> | | | 1 | -------------> | 1 |
> | | +-----------+ +-----------+
> | | | 2 | -------------> | 2 |
> | | +-----------+ +-----------+
> | | | 3 | -------------> | 3 |
> | | +-----------+ +-----------+
> | | | 4 | -------------> | 4 |
> | 2MB | +-----------+ +-----------+
> | | | 5 | -------------> | 5 |
> | | +-----------+ +-----------+
> | | | 6 | -------------> | 6 |
> | | +-----------+ +-----------+
> | | | 7 | -------------> | 7 |
> | | +-----------+ +-----------+
> | |
> | |
> | |
> +-----------+
>
> The value of page->compound_head is the same for all tail pages. The first
> page of page structs (page 0) associated with the HugeTLB page contains the 4
> page structs necessary to describe the HugeTLB. The only use of the remaining
> pages of page structs (page 1 to page 7) is to point to page->compound_head.
> Therefore, we can remap pages 2 to 7 to page 1. Only 2 pages of page structs
> will be used for each HugeTLB page. This will allow us to free the remaining
> 6 pages to the buddy allocator.
>
> Here is how things look after remapping.
>
> HugeTLB struct pages(8 pages) page frame(8 pages)
> +-----------+ ---virt_to_page---> +-----------+ mapping to +-----------+
> | | | 0 | -------------> | 0 |
> | | +-----------+ +-----------+
> | | | 1 | -------------> | 1 |
> | | +-----------+ +-----------+
> | | | 2 | ----------------^ ^ ^ ^ ^ ^
> | | +-----------+ | | | | |
> | | | 3 | ------------------+ | | | |
> | | +-----------+ | | | |
> | | | 4 | --------------------+ | | |
> | 2MB | +-----------+ | | |
> | | | 5 | ----------------------+ | |
> | | +-----------+ | |
> | | | 6 | ------------------------+ |
> | | +-----------+ |
> | | | 7 | --------------------------+
> | | +-----------+
> | |
> | |
> | |
> +-----------+
>
> When a HugeTLB is freed to the buddy system, we should allocate 6 pages for
> vmemmap pages and restore the previous mapping relationship.
>
> Apart from 2MB HugeTLB page, we also have 1GB HugeTLB page. It is similar
> to the 2MB HugeTLB page. We also can use this approach to free the vmemmap
> pages.
>
> In this case, for the 1GB HugeTLB page, we can save 4088 pages(There are
> 4096 pages for struct page structs, we reserve 2 pages for vmemmap and 8
> pages for page tables. So we can save 4088 pages). This is a very substantial
> gain. On our server, run some SPDK/QEMU applications which will use 1024GB
> hugetlbpage. With this feature enabled, we can save ~16GB(1G hugepage)/~11GB
> (2MB hugepage, the worst case is 10GB while the best is 12GB) memory.
>
> Because there are vmemmap page tables reconstruction on the freeing/allocating
> path, it increases some overhead. Here are some overhead analysis.
>
> 1) Allocating 10240 2MB hugetlb pages.
>
> a) With this patch series applied:
> # time echo 10240 > /proc/sys/vm/nr_hugepages
>
> real 0m0.166s
> user 0m0.000s
> sys 0m0.166s
>
> # bpftrace -e 'kprobe:alloc_fresh_huge_page { @start[tid] = nsecs; } kretprobe:alloc_fresh_huge_page /@start[tid]/ { @latency = hist(nsecs - @start[tid]); delete(@start[tid]); }'
> Attaching 2 probes...
>
> @latency:
> [8K, 16K) 8360 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@|
> [16K, 32K) 1868 |@@@@@@@@@@@ |
> [32K, 64K) 10 | |
> [64K, 128K) 2 | |
>
> b) Without this patch series:
> # time echo 10240 > /proc/sys/vm/nr_hugepages
>
> real 0m0.066s
> user 0m0.000s
> sys 0m0.066s
>
> # bpftrace -e 'kprobe:alloc_fresh_huge_page { @start[tid] = nsecs; } kretprobe:alloc_fresh_huge_page /@start[tid]/ { @latency = hist(nsecs - @start[tid]); delete(@start[tid]); }'
> Attaching 2 probes...
>
> @latency:
> [4K, 8K) 10176 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@|
> [8K, 16K) 62 | |
> [16K, 32K) 2 | |
>
> Summarize: this feature is about ~2x slower than before.
>
> 2) Freeing 10240 2MB hugetlb pages.
>
> a) With this patch series applied:
> # time echo 0 > /proc/sys/vm/nr_hugepages
>
> real 0m0.004s
> user 0m0.000s
> sys 0m0.002s
>
> # bpftrace -e 'kprobe:__free_hugepage { @start[tid] = nsecs; } kretprobe:__free_hugepage /@start[tid]/ { @latency = hist(nsecs - @start[tid]); delete(@start[tid]); }'
> Attaching 2 probes...
>
> @latency:
> [16K, 32K) 10240 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@|
>
> b) Without this patch series:
> # time echo 0 > /proc/sys/vm/nr_hugepages
>
> real 0m0.077s
> user 0m0.001s
> sys 0m0.075s
>
> # bpftrace -e 'kprobe:__free_hugepage { @start[tid] = nsecs; } kretprobe:__free_hugepage /@start[tid]/ { @latency = hist(nsecs - @start[tid]); delete(@start[tid]); }'
> Attaching 2 probes...
>
> @latency:
> [4K, 8K) 9950 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@|
> [8K, 16K) 287 |@ |
> [16K, 32K) 3 | |
>
> Summarize: The overhead of __free_hugepage is about ~2-4x slower than before.
> But according to the allocation test above, I think that here is
> also ~2x slower than before.
>
> But why the 'real' time of patched is smaller than before? Because
> In this patch series, the freeing hugetlb is asynchronous(through
> kwoker).
>
> Although the overhead has increased, the overhead is not significant. Like Mike
> said, "However, remember that the majority of use cases create hugetlb pages at
> or shortly after boot time and add them to the pool. So, additional overhead is
> at pool creation time. There is no change to 'normal run time' operations of
> getting a page from or returning a page to the pool (think page fault/unmap)".
>

Just FYI, I'll be offline until first week of January. I'm planning on
reviewing when I'm back.


--
Thanks,

David / dhildenb