Reducing Memory Consumption (aset and generation) - Mailing list pgsql-hackers

HI hackers,

I thought it would be better to start a new thread to discuss.

While working with sorting patch, and read others threads,

I have some ideas to reduces memory consumption by aset and generation

memory modules.

I have done basic benchmarks, and it seems to improve performance.

I think it's really worth it, if it really is possible to reduce memory consumption.

Linux Ubuntu 64 bits
work_mem = 64MB

set max_parallel_workers_per_gather = 0;
create table t (a bigint not null, b bigint not null, c bigint not
null, d bigint not null, e bigint not null, f bigint not null);

insert into t select x,x,x,x,x,x from generate_Series(1,140247142) x; -- 10GB!
vacuum freeze t;

select * from t order by a offset 140247142;

HEAD:
postgres=# select * from t order by a offset 140247142;
 a | b | c | d | e | f
---+---+---+---+---+---
(0 rows)

work_mem=64MB
Time: 99603,544 ms (01:39,604)
Time: 94000,342 ms (01:34,000)

postgres=# set work_mem="64.2MB";
SET
Time: 0,210 ms
postgres=# select * from t order by a offset 140247142;
 a | b | c | d | e | f
---+---+---+---+---+---
(0 rows)

Time: 95306,254 ms (01:35,306)


PATCHED:
postgres=# explain analyze select * from t order by a offset 140247142;
 a | b | c | d | e | f
---+---+---+---+---+---
(0 rows)

work_mem=64MB
Time: 90946,482 ms (01:30,946)

postgres=# set work_mem="64.2MB";
SET
Time: 0,210 ms
postgres=# select * from t order by a offset 140247142;
 a | b | c | d | e | f
---+---+---+---+---+---
(0 rows)

Time: 91817,533 ms (01:31,818)

There is still room for further improvements, and at this point I need help.

Regarding the patches we have:

1) 001-aset-reduces-memory-consumption.patch

Reduces memory used by struct AllocBlockData by minus 8 bits,

reducing the total size to 32 bits, which leads to "fitting" two structs in a 64bit cache.

Move some stores to fields struct, for the order of declaration, within the structures.

Remove tests elog(ERROR, "could not find block containing chunk %p" and

elog(ERROR, "could not find block containing chunk %p", moving them to

MEMORY_CONTEXT_CHECKING context.

Since 8.2 versions, nobody complains about these tests.

But if is not acceptable, have the option (3) 003-aset-reduces-memory-consumption.patch

2) 002-generation-reduces-memory-consumption.patch

Reduces memory used by struct GenerationBlock, by minus 8 bits,

reducing the total size to 32 bits, which leads to "fitting" two structs in a 64bit cache.

Remove all references to the field *block* used by struct GenerationChunk,

enabling its removal! (not done yet).

What would take the final size to 16 bits, which leads to "fitting" four structs in a 64bit cache.

Unfortunately, everything works only for the size 24, see the (4).

Move some stores to fields struct, for the order of declaration, within the structures.

3) 003-aset-reduces-memory-consumption.patch

Same to the (1), but without remove the tests:

But at the cost of removing a one tiny part of the tests.

4) 004-generation-reduces-memory-consumption-BUG.patch

Same to the (2), but with BUG.

It only takes a few tweaks to completely remove the field block.

@@ -117,9 +116,9 @@ struct GenerationChunk
  /* this is zero in a free chunk */
  Size requested_size;
 
-#define GENERATIONCHUNK_RAWSIZE  (SIZEOF_SIZE_T * 2 + SIZEOF_VOID_P * 2)
+#define GENERATIONCHUNK_RAWSIZE  (SIZEOF_SIZE_T * 2 + SIZEOF_VOID_P)
 #else
-#define GENERATIONCHUNK_RAWSIZE  (SIZEOF_SIZE_T + SIZEOF_VOID_P * 2)
+#define GENERATIONCHUNK_RAWSIZE  (SIZEOF_SIZE_T + SIZEOF_VOID_P)
 #endif /* MEMORY_CONTEXT_CHECKING */
 
  /* ensure proper alignment by adding padding if needed */
@@ -127,7 +126,6 @@ struct GenerationChunk
  char padding[MAXIMUM_ALIGNOF - GENERATIONCHUNK_RAWSIZE % MAXIMUM_ALIGNOF];
 #endif
 
- GenerationBlock *block; /* block owning this chunk */
  GenerationContext *context; /* owning context, or NULL if freed chunk */
  /* there must not be any padding to reach a MAXALIGN boundary here! */
};

This fails with make check.

I couldn't figure out why it doesn't work with 16 bits (struct GenerationChunk).