Patch series "mm, swap: introduce swap table as swap cache (phase I)", v3.
This is the first phase of the bigger series implementing basic
infrastructures for the Swap Table idea proposed at the LSF/MM/BPF topic
"Integrate swap cache, swap maps with swap allocator" [1]. To give credit
where it is due, this is based on Chris Li's idea and a prototype of using
cluster size atomic arrays to implement swap cache.
This phase I contains 15 patches, introduces the swap table infrastructure
and uses it as the swap cache backend. By doing so, we have up to ~5-20%
performance gain in throughput, RPS or build time for benchmark and
workload tests. The speed up is due to less contention on the swap cache
access and shallower swap cache lookup path. The cluster size is much
finer-grained than the 64M address space split, which is removed in this
phase I. It also unifies and cleans up the swap code base.
Each swap cluster will dynamically allocate the swap table, which is an
atomic array to cover every swap slot in the cluster. It replaces the
swap cache backed by XArray. In phase I, the static allocated swap_map
still co-exists with the swap table. The memory usage is about the same
as the original on average. A few exception test cases show about 1%
higher in memory usage. In the following phases of the series, swap_map
will merge into the swap table without additional memory allocation. It
will result in net memory reduction compared to the original swap cache.
Testing has shown that phase I has a significant performance improvement
from 8c/1G ARM machine to 48c96t/128G x86_64 servers in many practical
workloads.
The full picture with a summary can be found at [2]. An older bigger
series of 28 patches is posted at [3].
vm-scability test:
==================
Test with:
usemem --init-time -O -y -x -n 31 1G (4G memcg, PMEM as swap)
Before: After:
System time: 219.12s 158.16s (-27.82%)
Sum Throughput: 4767.13 MB/s 6128.59 MB/s (+28.55%)
Single process Throughput: 150.21 MB/s 196.52 MB/s (+30.83%)
Free latency: 175047.58 us 131411.87 us (-24.92%)
usemem --init-time -O -y -x -n 32 1536M (16G memory, global pressure,
PMEM as swap)
Before: After:
System time: 356.16s 284.68s (-20.06%)
Sum Throughput: 4648.35 MB/s 5453.52 MB/s (+17.32%)
Single process Throughput: 141.63 MB/s 168.35 MB/s (+18.86%)
Free latency: 499907.71 us 484977.03 us (-2.99%)
This shows an improvement of more than 20% improvement in most readings.
Build kernel test:
==================
The following result matrix is from building kernel with defconfig on
tmpfs with ZSWAP / ZRAM, using different memory pressure and setups.
Measuring sys and real time in seconds, less is better (user time is
almost identical as expected):
-j<NR> / Mem | Sys before / after | Real before / after
Using 16G ZRAM with memcg limit:
6 / 192M | 9686 / 9472 -2.21% | 2130 / 2096 -1.59%
12 / 256M | 6610 / 6451 -2.41% | 827 / 812 -1.81%
24 / 384M | 5938 / 5701 -3.37% | 414 / 405 -2.17%
48 / 768M | 4696 / 4409 -6.11% | 188 / 182 -3.19%
With 64k folio:
24 / 512M | 4222 / 4162 -1.42% | 326 / 321 -1.53%
48 / 1G | 3688 / 3622 -1.79% | 151 / 149 -1.32%
With ZSWAP with 3G memcg (using higher limit due to kmem account):
48 / 3G | 603 / 581 -3.65% | 81 / 80 -1.23%
Testing extremely high global memory and schedule pressure: Using ZSWAP
with 32G NVMEs in a 48c VM that has 4G memory, no memcg limit, system
components take up about 1.5G already, using make -j48 to build defconfig:
Before: sys time: 2069.53s real time: 135.76s
After: sys time: 2021.13s (-2.34%) real time: 134.23s (-1.12%)
On another 48c 4G memory VM, using 16G ZRAM as swap, testing make
-j48 with same config:
Before: sys time: 1756.96s real time: 111.01s
After: sys time: 1715.90s (-2.34%) real time: 109.51s (-1.35%)
All cases are more or less faster, and no regression even under extremely
heavy global memory pressure.
Redis / Valkey bench:
=====================
The test machine is a ARM64 VM with 1536M memory 12 cores, Redis is set to
use 2500M memory, and ZRAM swap size is set to 5G:
Testing with:
redis-benchmark -r
2000000 -n
2000000 -d 1024 -c 12 -P 32 -t get
no BGSAVE with BGSAVE
Before: 487576.06 RPS 280016.02 RPS
After: 487541.76 RPS (-0.01%) 300155.32 RPS (+7.19%)
Testing with:
redis-benchmark -r
2500000 -n
2500000 -d 1024 -c 12 -P 32 -t get
no BGSAVE with BGSAVE
Before: 466789.59 RPS 281213.92 RPS
After: 466402.89 RPS (-0.08%) 298411.84 RPS (+6.12%)
With BGSAVE enabled, most Redis memory will have a swap count > 1 so swap
cache is heavily in use. We can see a about 6% performance gain. No
BGSAVE is very slightly slower (<0.1%) due to the higher memory pressure
of the co-existence of swap_map and swap table. This will be optimzed
into a net gain and up to 20% gain in BGSAVE case in the following phases.
HDD swap is also ~40% faster with usemem because we removed an old
contention workaround.
This patch (of 15):
Swap table is the new swap cache.
Link: https://lkml.kernel.org/r/20250910160833.3464-1-ryncsn@gmail.com
Link: https://lkml.kernel.org/r/20250910160833.3464-2-ryncsn@gmail.com
Link: https://lore.kernel.org/CAMgjq7BvQ0ZXvyLGp2YP96+i+6COCBBJCYmjXHGBnfisCAb8VA@mail.gmail.com
Link: https://github.com/ryncsn/linux/tree/kasong/devel/swap-table
Link: https://lore.kernel.org/linux-mm/20250514201729.48420-1-ryncsn@gmail.com/
Signed-off-by: Chris Li <chrisl@kernel.org>
Signed-off-by: Kairui Song <kasong@tencent.com>
Cc: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: Baoquan He <bhe@redhat.com>
Cc: Barry Song <baohua@kernel.org>
Cc: David Hildenbrand <david@redhat.com>
Cc: "Huang, Ying" <ying.huang@linux.alibaba.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Kemeng Shi <shikemeng@huaweicloud.com>
Cc: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Cc: Zi Yan <ziy@nvidia.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
page_table_check
remap_file_pages
split_page_table_lock
+ swap-table
transhuge
unevictable-lru
vmalloced-kernel-stacks
--- /dev/null
+.. SPDX-License-Identifier: GPL-2.0
+
+:Author: Chris Li <chrisl@kernel.org>, Kairui Song <kasong@tencent.com>
+
+==========
+Swap Table
+==========
+
+Swap table implements swap cache as a per-cluster swap cache value array.
+
+Swap Entry
+----------
+
+A swap entry contains the information required to serve the anonymous page
+fault.
+
+Swap entry is encoded as two parts: swap type and swap offset.
+
+The swap type indicates which swap device to use.
+The swap offset is the offset of the swap file to read the page data from.
+
+Swap Cache
+----------
+
+Swap cache is a map to look up folios using swap entry as the key. The result
+value can have three possible types depending on which stage of this swap entry
+was in.
+
+1. NULL: This swap entry is not used.
+
+2. folio: A folio has been allocated and bound to this swap entry. This is
+ the transient state of swap out or swap in. The folio data can be in
+ the folio or swap file, or both.
+
+3. shadow: The shadow contains the working set information of the swapped
+ out folio. This is the normal state for a swapped out page.
+
+Swap Table Internals
+--------------------
+
+The previous swap cache is implemented by XArray. The XArray is a tree
+structure. Each lookup will go through multiple nodes. Can we do better?
+
+Notice that most of the time when we look up the swap cache, we are either
+in a swap in or swap out path. We should already have the swap cluster,
+which contains the swap entry.
+
+If we have a per-cluster array to store swap cache value in the cluster.
+Swap cache lookup within the cluster can be a very simple array lookup.
+
+We give such a per-cluster swap cache value array a name: the swap table.
+
+Each swap cluster contains 512 entries, so a swap table stores one cluster
+worth of swap cache values, which is exactly one page. This is not
+coincidental because the cluster size is determined by the huge page size.
+The swap table is holding an array of pointers. The pointer has the same
+size as the PTE. The size of the swap table should match to the second
+last level of the page table page, exactly one page.
+
+With swap table, swap cache lookup can achieve great locality, simpler,
+and faster.
+
+Locking
+-------
+
+Swap table modification requires taking the cluster lock. If a folio
+is being added to or removed from the swap table, the folio must be
+locked prior to the cluster lock. After adding or removing is done, the
+folio shall be unlocked.
+
+Swap table lookup is protected by RCU and atomic read. If the lookup
+returns a folio, the user must lock the folio before use.
R: Chris Li <chrisl@kernel.org>
L: linux-mm@kvack.org
S: Maintained
+F: Documentation/mm/swap-table.rst
F: include/linux/swap.h
F: include/linux/swapfile.h
F: include/linux/swapops.h
projects / users / jedix / linux-maple.git / commitdiff