xfs: improve metadata I/O merging in the elevator
Subsystems (2)
PATCH 1/1 xfs: improve metadata I/O merging in the elevator
I had the patch below from Dave in my queue for a while, but previously couldn't really reproduce his numbers. After some discussions of the bio types I've reteseted it again and can see constant improvements when using cfq on my large array box with it (5-10% for the sequential create workloads), but still nothing on deadline. Given that people also want it for better marking in blktrace it might be time to put it in.
Comments?
Signature · 72 lines
From: Dave Chinner <[email protected]> Subject: xfs: improve metadata I/O merging in the elevator Change all async metadata buffers to use [READ|WRITE]_META I/O types so that the I/O doesn't get issued immediately. This allows merging of adjacent metadata requests but still prioritises them over bulk data. This shows a 10-15% improvement in sequential create speed of small files. Don't include the log buffers in this classification - leave them as sync types so they are issued immediately. Signed-off-by: Dave Chinner <[email protected]> Signed-off-by: Christoph Hellwig <[email protected]> Index: xfs/fs/xfs/linux-2.6/xfs_buf.c =================================================================== --- xfs.orig/fs/xfs/linux-2.6/xfs_buf.c 2009-11-12 17:10:19.852253847 +0100 +++ xfs/fs/xfs/linux-2.6/xfs_buf.c 2009-11-12 17:13:55.334003777 +0100 @@ -1177,10 +1177,14 @@ _xfs_buf_ioapply( if (bp->b_flags & XBF_ORDERED) { ASSERT(!(bp->b_flags & XBF_READ)); rw = WRITE_BARRIER; - } else if (bp->b_flags & _XBF_RUN_QUEUES) { + } else if (bp->b_flags & XBF_LOG_BUFFER) { ASSERT(!(bp->b_flags & XBF_READ_AHEAD)); bp->b_flags &= ~_XBF_RUN_QUEUES; rw = (bp->b_flags & XBF_WRITE) ? WRITE_SYNC : READ_SYNC; + } else if (bp->b_flags & _XBF_RUN_QUEUES) { + ASSERT(!(bp->b_flags & XBF_READ_AHEAD)); + bp->b_flags &= ~_XBF_RUN_QUEUES; + rw = (bp->b_flags & XBF_WRITE) ? WRITE : READ_META; } else { rw = (bp->b_flags & XBF_WRITE) ? WRITE : (bp->b_flags & XBF_READ_AHEAD) ? READA : READ; Index: xfs/fs/xfs/linux-2.6/xfs_buf.h =================================================================== --- xfs.orig/fs/xfs/linux-2.6/xfs_buf.h 2009-11-12 17:10:19.857278370 +0100 +++ xfs/fs/xfs/linux-2.6/xfs_buf.h 2009-11-12 17:13:55.334003777 +0100 @@ -55,6 +55,7 @@ typedef enum { XBF_FS_MANAGED = (1 << 8), /* filesystem controls freeing memory */ XBF_ORDERED = (1 << 11), /* use ordered writes */ XBF_READ_AHEAD = (1 << 12), /* asynchronous read-ahead */ + XBF_LOG_BUFFER = (1 << 13), /* this is a buffer used for the log */ /* flags used only as arguments to access routines */ XBF_LOCK = (1 << 14), /* lock requested */ Index: xfs/fs/xfs/xfs_log.c =================================================================== --- xfs.orig/fs/xfs/xfs_log.c 2009-11-12 17:10:20.267254560 +0100 +++ xfs/fs/xfs/xfs_log.c 2009-11-12 17:13:55.335004184 +0100 @@ -1524,6 +1524,7 @@ xlog_sync(xlog_t *log, XFS_BUF_ZEROFLAGS(bp); XFS_BUF_BUSY(bp); XFS_BUF_ASYNC(bp); + bp->b_flags |= XBF_LOG_BUFFER; /* * Do an ordered write for the log block. * Its unnecessary to flush the first split block in the log wrap case. @@ -1561,6 +1562,7 @@ xlog_sync(xlog_t *log, XFS_BUF_ZEROFLAGS(bp); XFS_BUF_BUSY(bp); XFS_BUF_ASYNC(bp); + bp->b_flags |= XBF_LOG_BUFFER; if (log->l_mp->m_flags & XFS_MOUNT_BARRIER) XFS_BUF_ORDERED(bp); dptr = XFS_BUF_PTR(bp); _______________________________________________ xfs mailing list [email protected] http://oss.sgi.com/mailman/listinfo/xfs
Quoted 8 lines · Christoph Hellwig
I had the patch below from Dave in my queue for a while, but previously couldn't really reproduce his numbers. After some discussions of the bio types I've reteseted it again and can see constant improvements when using cfq on my large array box with it (5-10% for the sequential create workloads), but still nothing on deadline. Given that people also want it for better marking in blktrace it might be time to put it in.
Comments?
Definitely should be done, but....
It looks like the patch you posted isn't quite doing what was intended - async write buffers are being classified as WRITE, not WRITE_META. That means we get more write combining in the elevator (performance increase) like with WRITE_META, but don't get the faster dispatch (latency reduction) by using the META queue to keep the metadata writeback separate from the bulk data writeback. That may be why deadline is not showing any improvement...
FWIW, the original patch here:
http://oss.sgi.com/archives/xfs/2008-01/msg00630.html
uses WRITE_META, but it looks like you've taken bits of this patch:
http://oss.sgi.com/archives/xfs/2008-01/msg00653.html
and added the log buffer marking to this patch and accidentally dropped the WRITE_META marking. i.e. this:
Review
Quoted 4 lines
(quoted diff) +4
-0+ } else if (bp->b_flags & _XBF_RUN_QUEUES) { + ASSERT(!(bp->b_flags & XBF_READ_AHEAD)); + bp->b_flags &= ~_XBF_RUN_QUEUES; + rw = (bp->b_flags & XBF_WRITE) ? WRITE : READ_META;
I think should be:
+ rw = (bp->b_flags & XBF_WRITE) ? WRITE_META : READ_META;
Cheers,
Dave.
Signature · 7 lines
Dave Chinner [email protected] _______________________________________________ xfs mailing list [email protected] http://oss.sgi.com/mailman/listinfo/xfs
Quoted 10 lines · Dave Chinner
Definitely should be done, but....
It looks like the patch you posted isn't quite doing what was intended - async write buffers are being classified as WRITE, not
WRITE_META. That means we get more write combining in the elevator (performance increase) like withWRITE_META, but don't get the faster dispatch (latency reduction) by using the META queue to keep the metadata writeback separate from the bulk data writeback. That may be why deadline is not showing any improvement...
Quoted 6 lines
FWIW, the original patch here:
http://oss.sgi.com/archives/xfs/2008-01/msg00630.html
uses
WRITE_META, but it looks like you've taken bits of this patch:
Indeed. I'l re-add the write-side markings and will re-bench and re-submit.
_______________________________________________
xfs mailing list
[email protected]
http://oss.sgi.com/mailman/listinfo/xfs
From: Dave Chinner <[email protected]>
Change all async metadata buffers to use [READ|WRITE]_META I/O types so that the I/O doesn't get issued immediately. This allows merging of adjacent metadata requests but still prioritises them over bulk data. This shows a 10-15% improvement in sequential create speed of small files.
Don't include the log buffers in this classification - leave them as sync types so they are issued immediately.
fs/xfs/linux-2.6/xfs_buf.c +5 -1
@@ -1177,10 +1177,14 @@ _xfs_buf_ioapply(if (bp->b_flags &XBF_ORDERED ) {ASSERT (!(bp->b_flags &XBF_READ )); rw =WRITE_BARRIER ; - }else if (bp->b_flags & _XBF_RUN_QUEUES) { + }else if (bp->b_flags &XBF_LOG_BUFFER ) {ASSERT (!(bp->b_flags &XBF_READ_AHEAD )); bp->b_flags &= ~_XBF_RUN_QUEUES; rw = (bp->b_flags &XBF_WRITE ) ?WRITE_SYNC :READ_SYNC ; + }else if (bp->b_flags & _XBF_RUN_QUEUES) { +ASSERT (!(bp->b_flags &XBF_READ_AHEAD )); + bp->b_flags &= ~_XBF_RUN_QUEUES; + rw = (bp->b_flags &XBF_WRITE ) ?WRITE_META :READ_META ; }else { rw = (bp->b_flags &XBF_WRITE ) ?WRITE : (bp->b_flags &XBF_READ_AHEAD ) ?READA :READ ;
fs/xfs/linux-2.6/xfs_buf.h +1 -0
@@ -55,6 +55,7 @@ typedef enum {XBF_FS_MANAGED = (1 <<8 ),/* filesystem controls freeing memory */ XBF_ORDERED = (1 <<11 ),/* use ordered writes */ XBF_READ_AHEAD = (1 <<12 ),/* asynchronous read-ahead */ +XBF_LOG_BUFFER = (1 <<13 ),/* this is a buffer used for the log */ /* flags used only as arguments to access routines */ XBF_LOCK = (1 << 14), /* lock requested */
fs/xfs/xfs_log.c +2 -0
@@ -1524,6 +1524,7 @@ xlog_sync(xlog_t *log,XFS_BUF_ZEROFLAGS (bp);XFS_BUF_BUSY (bp);XFS_BUF_ASYNC (bp); + bp->b_flags |=XBF_LOG_BUFFER ;/* * Do an ordered write for the log block. * Its unnecessary to flush the first split block in the log wrap case.
@@ -1561,6 +1562,7 @@ xlog_sync(xlog_t *log,XFS_BUF_ZEROFLAGS (bp);XFS_BUF_BUSY (bp);XFS_BUF_ASYNC (bp); + bp->b_flags |=XBF_LOG_BUFFER ;if (log->l_mp->m_flags &XFS_MOUNT_BARRIER )XFS_BUF_ORDERED (bp); dptr =XFS_BUF_PTR (bp);
include/linux/fs.h +1 -0
@@ -151,6 +151,7 @@ struct inodes_stat_t {#define READ_META (READ | (1 << BIO_RW_META))#define WRITE_SYNC_PLUG (WRITE | (1 << BIO_RW_SYNCIO) | (1 << BIO_RW_NOIDLE))#define WRITE_SYNC (WRITE_SYNC_PLUG | (1 << BIO_RW_UNPLUG)) +#define WRITE_META (WRITE | (1 << BIO_RW_META))#define WRITE_ODIRECT (WRITE | (1 << BIO_RW_SYNCIO) | (1 << BIO_RW_UNPLUG))#define SWRITE_SYNC_PLUG \ (SWRITE | (1 << BIO_RW_SYNCIO) | (1 << BIO_RW_NOIDLE))
_______________________________________________
xfs mailing list
[email protected]
http://oss.sgi.com/mailman/listinfo/xfs
Alex Elder
Reviewed-byQuoted 10 lines · Christoph Hellwig
From: Dave Chinner <[email protected]>
Change all async metadata buffers to use [READ|WRITE]_META I/O types so that the I/O doesn't get issued immediately. This allows merging of adjacent metadata requests but still prioritises them over bulk data. This shows a 10-15% improvement in sequential create speed of small files.
Don't include the log buffers in this classification - leave them as sync types so they are issued immediately.
Looks good.
Quoted 2 lines
Signed-off-by Dave Chinner <[email protected]>Signed-off-by Christoph Hellwig <[email protected]>
Quoted 65 lines
Index:
xfs/fs/xfs/linux-2.6/xfs_buf.c
===================================================================fs/xfs/linux-2.6/xfs_buf.c +5
-1@@ -1177,10 +1177,14 @@ _xfs_buf_ioapply(if (bp->b_flags &XBF_ORDERED ) {ASSERT (!(bp->b_flags &XBF_READ )); rw =WRITE_BARRIER ; - }else if (bp->b_flags & _XBF_RUN_QUEUES) { + }else if (bp->b_flags &XBF_LOG_BUFFER ) {ASSERT (!(bp->b_flags &XBF_READ_AHEAD )); bp->b_flags &= ~_XBF_RUN_QUEUES; rw = (bp->b_flags &XBF_WRITE ) ?WRITE_SYNC :READ_SYNC ; + }else if (bp->b_flags & _XBF_RUN_QUEUES) { +ASSERT (!(bp->b_flags &XBF_READ_AHEAD )); + bp->b_flags &= ~_XBF_RUN_QUEUES; + rw = (bp->b_flags &XBF_WRITE ) ?WRITE_META :READ_META ; }else { rw = (bp->b_flags &XBF_WRITE ) ?WRITE : (bp->b_flags &XBF_READ_AHEAD ) ?READA :READ ;Index:
xfs/fs/xfs/linux-2.6/xfs_buf.h
===================================================================fs/xfs/linux-2.6/xfs_buf.h +1
-0@@ -55,6 +55,7 @@ typedef enum {XBF_FS_MANAGED = (1 <<8 ),/* filesystem controls freeing memory */ XBF_ORDERED = (1 <<11 ),/* use ordered writes */ XBF_READ_AHEAD = (1 <<12 ),/* asynchronous read-ahead */ +XBF_LOG_BUFFER = (1 <<13 ),/* this is a buffer used for the log */ /* flags used only as arguments to access routines */ XBF_LOCK = (1 << 14), /* lock requested */Index:
xfs/fs/xfs/xfs_log.c
===================================================================fs/xfs/xfs_log.c +2
-0@@ -1524,6 +1524,7 @@ xlog_sync(xlog_t *log,XFS_BUF_ZEROFLAGS (bp);XFS_BUF_BUSY (bp);XFS_BUF_ASYNC (bp); + bp->b_flags |=XBF_LOG_BUFFER ;/* * Do an ordered write for the log block. * Its unnecessary to flush the first split block in the log wrap case. @@ -1561,6 +1562,7 @@ xlog_sync(xlog_t *log,XFS_BUF_ZEROFLAGS (bp);XFS_BUF_BUSY (bp);XFS_BUF_ASYNC (bp); + bp->b_flags |=XBF_LOG_BUFFER ;if (log->l_mp->m_flags &XFS_MOUNT_BARRIER )XFS_BUF_ORDERED (bp); dptr =XFS_BUF_PTR (bp);Index:
xfs/include/linux/fs.h
===================================================================include/linux/fs.h +1
-0@@ -151,6 +151,7 @@ struct inodes_stat_t {#define READ_META (READ | (1 << BIO_RW_META))#define WRITE_SYNC_PLUG (WRITE | (1 << BIO_RW_SYNCIO) | (1 << BIO_RW_NOIDLE))#define WRITE_SYNC (WRITE_SYNC_PLUG | (1 << BIO_RW_UNPLUG)) +#define WRITE_META (WRITE | (1 << BIO_RW_META))#define WRITE_ODIRECT (WRITE | (1 << BIO_RW_SYNCIO) | (1 << BIO_RW_UNPLUG))#define SWRITE_SYNC_PLUG \ (SWRITE | (1 << BIO_RW_SYNCIO) | (1 << BIO_RW_NOIDLE))
Quoted 4 lines
_______________________________________________
xfs mailing list
[email protected]
http://oss.sgi.com/mailman/listinfo/xfs
_______________________________________________
xfs mailing list
[email protected]
http://oss.sgi.com/mailman/listinfo/xfs