Обсуждение: Align large shared memory allocations
Attached is a patch that aligns large shared memory allocations beyond
MAXIMUM_ALIGNOF. The reason for this is that Intel's cpus have a fast
path for bulk memory copies that only works with aligned addresses. It's
possible that other cpus have similar restrictions.
With 7.3.4, it achives a 5% performance gain with pgbench. It has no
effect with 7.3.3, because the buffers are already aligned by chance. I
haven't properly tested 7.4cvs yet.
One problem is the "32" - it's arbitrary, it probably belongs into an
arch dependant header file. But where?
--
Manfred
diff -u pgsql.orig/src/backend/storage/ipc/shmem.c pgsql/src/backend/storage/ipc/shmem.c
--- pgsql.orig/src/backend/storage/ipc/shmem.c 2003-09-20 20:17:08.000000000 +0200
+++ pgsql/src/backend/storage/ipc/shmem.c 2003-09-20 20:34:21.000000000 +0200
@@ -131,6 +131,7 @@
void *
ShmemAlloc(Size size)
{
+ uint32 newStart;
uint32 newFree;
void *newSpace;
@@ -146,10 +147,21 @@
SpinLockAcquire(ShmemLock);
- newFree = shmemseghdr->freeoffset + size;
+ newStart = shmemseghdr->freeoffset;
+ if (size >= BLCKSZ)
+ {
+ /* Align BLCKSZ sized buffers even further:
+ * - the costs are small
+ * - some cpus (most notably Intel Pentium III)
+ * prefer well-aligned addresses for memory copies
+ */
+ newStart = TYPEALIGN(32, newStart);
+ }
+
+ newFree = newStart + size;
if (newFree <= shmemseghdr->totalsize)
{
- newSpace = (void *) MAKE_PTR(shmemseghdr->freeoffset);
+ newSpace = (void *) MAKE_PTR(newStart);
shmemseghdr->freeoffset = newFree;
}
else
Manfred Spraul <manfred@colorfullife.com> writes:
> Attached is a patch that aligns large shared memory allocations beyond
> MAXIMUM_ALIGNOF. The reason for this is that Intel's cpus have a fast
> path for bulk memory copies that only works with aligned addresses.
This patch is missing a demonstration that it's actually worth anything.
What kind of performance gain do you get?
> One problem is the "32" - it's arbitrary, it probably belongs into an
> arch dependant header file. But where?
We don't really have arch-dependent header files. What I'd be inclined
to do is "#define ALIGNOF_BUFFER 32" in pg_config_manual.h, then
#define BUFFERALIGN(LEN) to parallel the other TYPEALIGN macros in c.h,
and finally use that in the ShmemAlloc code.
regards, tom lane
Tom Lane wrote:
>Manfred Spraul <manfred@colorfullife.com> writes:
>
>
>>Attached is a patch that aligns large shared memory allocations beyond
>>MAXIMUM_ALIGNOF. The reason for this is that Intel's cpus have a fast
>>path for bulk memory copies that only works with aligned addresses.
>>
>>
>
>This patch is missing a demonstration that it's actually worth anything.
>What kind of performance gain do you get?
>
>
7.4cvs on a 1.13 GHz Intel Celeron mobile, 384 MB RAM, "Severn" RedHat
Linux 2.4 beta, postmaster -N 30 -B 64, data directory on ramdisk,
pgbench -c 10 -s 11 -t 1000:
Without the patch: 124 tps
with the patch: 130 tps.
I've reduced the buffer setting to 64 because without that, a too large
part of the database was cached by postgres. I expect that with all
Intel Pentium III chips, it will be worth 10-20% less system time. I had
around 30% system time after reducing the number of buffers, thus the
~5% performance improvement.
>We don't really have arch-dependent header files. What I'd be inclined
>to do is "#define ALIGNOF_BUFFER 32" in pg_config_manual.h, then
>#define BUFFERALIGN(LEN) to parallel the other TYPEALIGN macros in c.h,
>and finally use that in the ShmemAlloc code.
>
>
Ok, new patch attached.
--
Manfred
Index: src/backend/storage/ipc/shmem.c
===================================================================
RCS file: /projects/cvsroot/pgsql-server/src/backend/storage/ipc/shmem.c,v
retrieving revision 1.70
diff -u -r1.70 shmem.c
--- src/backend/storage/ipc/shmem.c 4 Aug 2003 02:40:03 -0000 1.70
+++ src/backend/storage/ipc/shmem.c 21 Sep 2003 07:53:13 -0000
@@ -131,6 +131,7 @@
void *
ShmemAlloc(Size size)
{
+ uint32 newStart;
uint32 newFree;
void *newSpace;
@@ -146,10 +147,14 @@
SpinLockAcquire(ShmemLock);
- newFree = shmemseghdr->freeoffset + size;
+ newStart = shmemseghdr->freeoffset;
+ if (size >= BLCKSZ)
+ newStart = BUFFERALIGN(newStart);
+
+ newFree = newStart + size;
if (newFree <= shmemseghdr->totalsize)
{
- newSpace = (void *) MAKE_PTR(shmemseghdr->freeoffset);
+ newSpace = (void *) MAKE_PTR(newStart);
shmemseghdr->freeoffset = newFree;
}
else
Index: src/include/c.h
===================================================================
RCS file: /projects/cvsroot/pgsql-server/src/include/c.h,v
retrieving revision 1.152
diff -u -r1.152 c.h
--- src/include/c.h 4 Aug 2003 02:40:10 -0000 1.152
+++ src/include/c.h 21 Sep 2003 07:53:14 -0000
@@ -529,6 +529,7 @@
#define LONGALIGN(LEN) TYPEALIGN(ALIGNOF_LONG, (LEN))
#define DOUBLEALIGN(LEN) TYPEALIGN(ALIGNOF_DOUBLE, (LEN))
#define MAXALIGN(LEN) TYPEALIGN(MAXIMUM_ALIGNOF, (LEN))
+#define BUFFERALIGN(LEN) TYPEALIGN(ALIGNOF_BUFFER, (LEN))
/* ----------------------------------------------------------------
Index: src/include/pg_config_manual.h
===================================================================
RCS file: /projects/cvsroot/pgsql-server/src/include/pg_config_manual.h,v
retrieving revision 1.5
diff -u -r1.5 pg_config_manual.h
--- src/include/pg_config_manual.h 4 Aug 2003 00:43:29 -0000 1.5
+++ src/include/pg_config_manual.h 21 Sep 2003 07:53:14 -0000
@@ -176,6 +176,14 @@
*/
#define MAX_RANDOM_VALUE (0x7FFFFFFF)
+/*
+ * Alignment of the disk blocks in the shared memory area.
+ * A significant amount of the total system time is required for
+ * copying disk blocks between the os buffers and the cache in the
+ * shared memory area. Some cpus (most notably Intel Pentium III)
+ * prefer well-aligned addresses for memory copies.
+ */
+#define ALIGNOF_BUFFER 32
/*
*------------------------------------------------------------------------
Manfred Spraul <manfred@colorfullife.com> writes:
> Tom Lane wrote:
>> This patch is missing a demonstration that it's actually worth anything.
>> What kind of performance gain do you get?
>>
> 7.4cvs on a 1.13 GHz Intel Celeron mobile, 384 MB RAM, "Severn" RedHat
> Linux 2.4 beta, postmaster -N 30 -B 64, data directory on ramdisk,
> pgbench -c 10 -s 11 -t 1000:
> Without the patch: 124 tps
> with the patch: 130 tps.
I tried it on an Intel box here (P4 I think). Using postmaster -B 64 -N 30
and three tries of pgbench -s 10 -c 1 -t 1000 after creation of the test
tables, I get:
tps = 92.461144 (including connections establishing)
tps = 92.500572 (excluding connections establishing)
tps = 88.078814 (including connections establishing)
tps = 88.115905 (excluding connections establishing)
tps = 85.434473 (including connections establishing)
tps = 85.468807 (excluding connections establishing)
and with the patch:
tps = 122.927066 (including connections establishing)
tps = 122.998129 (excluding connections establishing)
tps = 110.716370 (including connections establishing)
tps = 110.773928 (excluding connections establishing)
tps = 138.155991 (including connections establishing)
tps = 138.245777 (excluding connections establishing)
So there's definitely a visible difference on recent Pentiums. It might
not help on other CPUs, but we can surely waste a couple dozen bytes in
the hope that it might.
Patch applied. Do you want to look at making it happen for local
buffers and buffile.c as well?
regards, tom lane