Hi list, PostgreSQL's default settings change when built with Linux kernel headers 2.6.33 or newer. As discussed on the pgsql-performance list, this causes a significant performance regression: http://archives.postgresql.org/pgsql-performance/2010-10/msg00602.php
NB! I am not proposing to change the default -- to the contrary -- this patch restores old behavior. Users might be in for a nasty performance surprise when re-building their Postgres with newer Linux headers (as was I), so I propose that this change should be made in all supported releases. -- commit message -- Revert default wal_sync_method to fdatasync on Linux 2.6.33+ Linux kernel headers from 2.6.33 (and later) change the behavior of the O_SYNC flag. Previously O_SYNC was aliased to O_DSYNC, which caused PostgreSQL to use fdatasync as the default instead. Starting with kernels 2.6.33 and later, the definitions of O_DSYNC and O_SYNC differ. When built with headers from these newer kernels, PostgreSQL will default to using open_datasync. This patch reverts the Linux default to fdatasync, which has had much more testing over time and also significantly better performance. -- end commit message -- Earlier kernel headers defined O_SYNC and O_DSYNC to 0x1000 2.6.33 and later define O_SYNC=0x101000 and O_DSYNC=0x1000 (since old behavior on most FS-es was always equivalent to POSIX O_DSYNC) More details at: http://git.kernel.org/?p=linux/kernel/git/torvalds/linux-2.6.git;a=commit;h=6b2f3d1f769be5779b479c37800229d9a4809fc3 Currently PostgreSQL's include/access/xlogdefs.h defaults to using open_datasync when O_SYNC != O_DSYNC, otherwise fdatasync is used. Since other platforms might want to default to fdatasync in the future, too, I defined a new PLATFORM_DEFAULT_SYNC_METHOD constant in include/port/linux.h. I don't know if this is the best way to do it. Regards, Marti
From fbf61c6536b7060e5c6745c8221a5a4fb9a53c92 Mon Sep 17 00:00:00 2001 From: Marti Raudsepp <ma...@juffo.org> Date: Fri, 5 Nov 2010 17:40:22 +0200 Subject: [PATCH] Revert default wal_sync_method to fdatasync on Linux 2.6.33+ Linux kernel headers from 2.6.33 (and later) change the behavior of the O_SYNC flag. Previously O_SYNC was aliased to O_DSYNC, which caused PostgreSQL to use fdatasync as the default instead. Starting with kernels 2.6.33 and later, the definitions of O_DSYNC and O_SYNC differ. When built with headers from these newer kernels, PostgreSQL will default to using open_datasync. This patch reverts the Linux default to fdatasync, which has had much more testing over time and also significantly better performance. --- src/include/access/xlogdefs.h | 5 +++++ src/include/port/linux.h | 9 +++++++++ 2 files changed, 14 insertions(+), 0 deletions(-) diff --git a/src/include/access/xlogdefs.h b/src/include/access/xlogdefs.h index 18b214e..7ed1bde 100644 --- a/src/include/access/xlogdefs.h +++ b/src/include/access/xlogdefs.h @@ -13,6 +13,7 @@ #define XLOG_DEFS_H #include <fcntl.h> /* need open() flags */ +#include "pg_config_os.h" /* * Pointer to a location in the XLOG. These pointers are 64 bits wide, @@ -123,6 +124,9 @@ typedef uint32 TimeLineID; #endif #endif +#if defined(PLATFORM_DEFAULT_SYNC_METHOD) +#define DEFAULT_SYNC_METHOD PLATFORM_DEFAULT_SYNC_METHOD +#else /* !defined(PLATFORM_DEFAULT_SYNC_METHOD) */ #if defined(OPEN_DATASYNC_FLAG) #define DEFAULT_SYNC_METHOD SYNC_METHOD_OPEN_DSYNC #elif defined(HAVE_FDATASYNC) @@ -132,6 +136,7 @@ typedef uint32 TimeLineID; #else #define DEFAULT_SYNC_METHOD SYNC_METHOD_FSYNC #endif +#endif /* * Limitation of buffer-alignment for direct IO depends on OS and filesystem, diff --git a/src/include/port/linux.h b/src/include/port/linux.h index b9498b2..5a88590 100644 --- a/src/include/port/linux.h +++ b/src/include/port/linux.h @@ -12,3 +12,12 @@ * to have a kernel version test here. */ #define HAVE_LINUX_EIDRM_BUG + +/* + * The definition of O_SYNC changed in Linux kernel headers starting with + * 2.6.33. This caused PostgreSQL to change the wal_sync_method default from + * fdatasync to open_datasync. Since fdatasync is more tested on Linux and has + * better performance, default to fdatasync on Linux. + */ +#define PLATFORM_DEFAULT_SYNC_METHOD SYNC_METHOD_FDATASYNC + -- 1.7.3.2
-- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
