On Wed, Jul 24, 2024 at 2:42 PM John Naylor <johncnaylo...@gmail.com> wrote:
> As for lowering the limit, we've experimented with 256kB here:
>
> https://www.postgresql.org/message-id/canwcazzutvz3lsypauyqvzcezxz7qe+9ntnhgyzdtwxpul+...@mail.gmail.com
>
> As I mention there, going lower than that would need a small amount of
> reorganization in the radix tree. Not difficult -- the thing I'm
> concerned about is that we'd likely need to document a separate
> minimum for DSA, since that behaves strangely with 256kB and might not
> work at all lower than that.
For experimentation, here's a rough patch (really two, squashed
together for now) that allows m_w_m to go down to 64kB.
drop table if exists test;
create table test (a int) with (autovacuum_enabled=false, fillfactor=10);
insert into test (a) select i from generate_series(1,2000) i;
create index on test (a);
update test set a = a + 1;
set maintenance_work_mem = '64kB';
vacuum (verbose) test;
INFO: vacuuming "john.public.test"
INFO: finished vacuuming "john.public.test": index scans: 3
pages: 0 removed, 91 remain, 91 scanned (100.00% of total)
The advantage with this is that we don't need to care about
MEMORY_CONTEXT_CHECKING or 32/64 bit-ness, since allocating a single
large node will immediately blow the limit, and that will happen
fairly quickly regardless. I suspect going this low will not work with
dynamic shared memory and if so would need a warning comment.
From 25ccfb9ed812c72194bd585906f9f11e18400422 Mon Sep 17 00:00:00 2001
From: John Naylor <john.nay...@postgresql.org>
Date: Wed, 24 Jul 2024 18:41:51 +0700
Subject: [PATCH v1] WIP: set minimum maintenance_work_mem to 64kB, matching
work_mem
To allow this to work for vacuum, create radix tree node
contexts lazily.
---
src/backend/utils/misc/guc_tables.c | 2 +-
src/include/lib/radixtree.h | 44 +++++++++++++++++++++--------
2 files changed, 34 insertions(+), 12 deletions(-)
diff --git a/src/backend/utils/misc/guc_tables.c b/src/backend/utils/misc/guc_tables.c
index f6fcdebb03..853ad8dba1 100644
--- a/src/backend/utils/misc/guc_tables.c
+++ b/src/backend/utils/misc/guc_tables.c
@@ -2445,7 +2445,7 @@ struct config_int ConfigureNamesInt[] =
GUC_UNIT_KB
},
&maintenance_work_mem,
- 65536, 1024, MAX_KILOBYTES,
+ 65536, 64, MAX_KILOBYTES,
NULL, NULL, NULL
},
diff --git a/src/include/lib/radixtree.h b/src/include/lib/radixtree.h
index 88bf695e3f..bd83b670dd 100644
--- a/src/include/lib/radixtree.h
+++ b/src/include/lib/radixtree.h
@@ -834,12 +834,30 @@ RT_ALLOC_NODE(RT_RADIX_TREE * tree, const uint8 kind, const RT_SIZE_CLASS size_c
RT_CHILD_PTR allocnode;
RT_NODE *node;
size_t allocsize;
+ RT_SIZE_CLASS_ELEM class_info;
- allocsize = RT_SIZE_CLASS_INFO[size_class].allocsize;
+ class_info = RT_SIZE_CLASS_INFO[size_class];
+ allocsize = class_info.allocsize;
#ifdef RT_SHMEM
allocnode.alloc = dsa_allocate(tree->dsa, allocsize);
#else
+
+ /*
+ * The context for RT_CLASS_4 should have been created when the tree was
+ * initialized.
+ */
+ Assert(tree->node_slabs[RT_CLASS_4] != NULL);
+ if (size_class > RT_CLASS_4 && tree->node_slabs[size_class] == NULL)
+ {
+ size_t blocksize = RT_SLAB_BLOCK_SIZE(allocsize);
+
+ tree->node_slabs[size_class] = SlabContextCreate(tree->context,
+ class_info.name,
+ blocksize,
+ allocsize);
+ }
+
allocnode.alloc = (RT_PTR_ALLOC) MemoryContextAlloc(tree->node_slabs[size_class],
allocsize);
#endif
@@ -1827,6 +1845,9 @@ RT_CREATE(MemoryContext ctx)
RT_CHILD_PTR rootnode;
#ifdef RT_SHMEM
dsa_pointer dp;
+#else
+ RT_SIZE_CLASS_ELEM class_info;
+ size_t node_blocksize;
#endif
old_ctx = MemoryContextSwitchTo(ctx);
@@ -1852,17 +1873,18 @@ RT_CREATE(MemoryContext ctx)
#else
tree->ctl = (RT_RADIX_TREE_CONTROL *) palloc0(sizeof(RT_RADIX_TREE_CONTROL));
- /* Create a slab context for each size class */
- for (int i = 0; i < RT_NUM_SIZE_CLASSES; i++)
- {
- RT_SIZE_CLASS_ELEM size_class = RT_SIZE_CLASS_INFO[i];
- size_t inner_blocksize = RT_SLAB_BLOCK_SIZE(size_class.allocsize);
+ /*
+ * For now only create a slab context for the smallest size class. The
+ * rest will be created on demand. This allows us to stay within memory
+ * settings that are smaller than the size of the larger slab blocks.
+ */
+ class_info = RT_SIZE_CLASS_INFO[RT_CLASS_4];
+ node_blocksize = RT_SLAB_BLOCK_SIZE(class_info.allocsize);
- tree->node_slabs[i] = SlabContextCreate(ctx,
- size_class.name,
- inner_blocksize,
- size_class.allocsize);
- }
+ tree->node_slabs[RT_CLASS_4] = SlabContextCreate(ctx,
+ class_info.name,
+ node_blocksize,
+ class_info.allocsize);
/* By default we use the passed context for leaves. */
tree->leaf_context = tree->context;
--
2.45.2
