Skip to contentSkip to conversation

Serious performance degradation on a RAID with kernel 2.6.10-bk7 and later

PATCH Andreas Hirstius 5 participants 10 messages Apr 20, 2005 – Apr 21, 2005 1 message not in the archive

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

Hi,

We have a rx4640 with 3x 3Ware 9500 SATA controllers and 24x WD740GD HDD in a software RAID0 configuration (using md). With kernel 2.6.11 the read performance on the md is reduced by a factor of 20 (!!) compared to previous kernels. The write rate to the md doesn't change!! (it actually improves a bit).

The config for the kernels are basically identical.

Here is some vmstat output:

kernel 2.6.9: ~1GB/s read
procs memory swap io system cpu

r  b   swpd   free   buff  cache   si   so    bi    bo   in    cs us sy wa id
1  1      0  12672   6592 15914112    0    0 1081344    56 15719  1583 0 11 14 74
1  0      0  12672   6592 15915200    0    0 1130496     0 15996  1626 0 11 14 74
0  1      0  12672   6592 15914112    0    0 1081344     0 15891  1570 0 11 14 74
0  1      0  12480   6592 15914112    0    0 1081344     0 15855  1537 0 11 14 74
1  0      0  12416   6592 15914112    0    0 1130496     0 16006  1586 0 12 14 74

kernel 2.6.11: ~55MB/s read
procs memory swap io system cpu

r  b   swpd   free   buff  cache   si   so    bi    bo   in    cs us sy wa id
1  1      0  24448  37568 15905984    0    0 56934     0 5166  1862  0 1 24 75
0  1      0  20672  37568 15909248    0    0 57280     0 5168  1871  0 1 24 75
0  1      0  22848  37568 15907072    0    0 57306     0 5173  1874  0 1 24 75
0  1      0  25664  37568 15903808    0    0 57190     0 5171  1870  0 1 24 75
0  1      0  21952  37568 15908160    0    0 57267     0 5168  1871  0 1 24 75

Because the filesystem might have an impact on the measurement, "dd" on /dev/md0
was used to get information about the performance. This also opens the possibility to test with block sizes larger than the page size. And it appears that the performance with kernel 2.6.11 is closely related to the block size.
For example if the block size is exactly a multiple (>2) of the page size the performance is back to ~1.1GB/s. The general behaviour is a bit more complicated:

1. bs <= 1.5 * ps : ~27-57MB/s (differs with ps)
2. bs > 1.5 * ps && bs < 2 * ps : rate increases to max. rate
3. bs = n * ps ; (n >= 2) : ~1.1GB/s (== max. rate)
4. bs > n * ps && bs < ~(n+0.5) * ps ; (n > 2) : ~27-70MB/s (differs
with ps)
5. bs > ~(n+0.5) * ps && bs < (n+1) * ps ; (n > 2) : increasing rate
in several, more or
less, distinct steps (e.g. 1/3 of max. rate and then 2/3 of max
rate for 64k pages)

I've tested all four possible page sizes on Itanium (4k, 8k, 16k and 64k) and the pattern is always the same!!

With kernel 2.6.9 (any kernel before 2.6.10-bk6) the read rate is always at ~1.1GB/s, independent of the block size.

This simple patch solves the problem, but I have no idea of possible side-effects ...

mm/filemap.c 2005-04-20 10:27:42.000000000 +0200 renamed from mm/filemap.c 2005-04-04 18:40:05.000000000 +0200 +1 -1
@@ -719,7 +719,7 @@
       index = *ppos >> PAGE_CACHE_SHIFT;
       next_index = index;
       prev_index = ra.prev_page;
-       last_index = (*ppos + desc->count + PAGE_CACHE_SIZE-1) >> PAGE_CACHE_SHIFT;
+       last_index = (*ppos + desc->count + PAGE_CACHE_SIZE) >> PAGE_CACHE_SHIFT;
       offset = *ppos & ~PAGE_CACHE_MASK;
 
       isize = i_size_read(inode);
mm/readahead.c 2005-04-20 18:37:04.000000000 +0200 renamed from mm/readahead.c 2005-04-04 18:40:05.000000000 +0200 +0 -0
@@ -70,7 +70,7 @@
 */

static unsigned long get_init_ra_size(unsigned long size, unsigned long max) {
- unsigned long newsize = roundup_pow_of_two(size);
+ unsigned long newsize = size;

       if (newsize <= max / 64)
               newsize = newsize * newsize;

In order to keep this mail short, I've created a webpage that contains all the detailed information and some plots:
http://www.cern.ch/openlab-debugging/raid

Regards,

Andreas Hirstius

#1 lore raw

For 3Ware, you need to chage the queue depths, and you will see dramatically improved performance. 3Ware can take requests a lot faster than Linux pushes them out. Try changing this instead, you won't be going to sleep all the time waiting on the read/write request queues to get "unstarved".

/linux/include/linux/blkdev.h

//#define BLKDEV_MIN_RQ 4
//#define BLKDEV_MAX_RQ 128 /* Default maximum */
#define BLKDEV_MIN_RQ 4096
#define BLKDEV_MAX_RQ 8192 /* Default maximum */

Jeff

Quoted 95 lines · Andreas Hirstius

Hi,

We have a rx4640 with 3x 3Ware 9500 SATA controllers and 24x WD740GD HDD in a software RAID0 configuration (using md). With kernel 2.6.11 the read performance on the md is reduced by a factor of 20 (!!) compared to previous kernels. The write rate to the md doesn't change!! (it actually improves a bit).

The config for the kernels are basically identical.

Here is some vmstat output:

kernel 2.6.9: ~1GB/s read
procs memory swap io system cpu
r b swpd free buff cache si so bi bo in cs us sy wa id 1 1 0 12672 6592 15914112 0 0 1081344 56 15719 1583 0 11 14 74 1 0 0 12672 6592 15915200 0 0 1130496 0 15996 1626 0 11 14 74 0 1 0 12672 6592 15914112 0 0 1081344 0 15891 1570 0 11 14 74 0 1 0 12480 6592 15914112 0 0 1081344 0 15855 1537 0 11 14 74 1 0 0 12416 6592 15914112 0 0 1130496 0 16006 1586 0 12 14 74

kernel 2.6.11: ~55MB/s read
procs memory swap io system cpu
r b swpd free buff cache si so bi bo in cs us sy wa id 1 1 0 24448 37568 15905984 0 0 56934 0 5166 1862 0 1 24 75 0 1 0 20672 37568 15909248 0 0 57280 0 5168 1871 0 1 24 75 0 1 0 22848 37568 15907072 0 0 57306 0 5173 1874 0 1 24 75 0 1 0 25664 37568 15903808 0 0 57190 0 5171 1870 0 1 24 75 0 1 0 21952 37568 15908160 0 0 57267 0 5168 1871 0 1 24 75

Because the filesystem might have an impact on the measurement, "dd" on /dev/md0
was used to get information about the performance. This also opens the possibility to test with block sizes larger than the page size. And it appears that the performance with kernel 2.6.11 is closely related to the block size.
For example if the block size is exactly a multiple (>2) of the page size the performance is back to ~1.1GB/s. The general behaviour is a bit more complicated:
1. bs <= 1.5 * ps : ~27-57MB/s (differs with ps)
2. bs > 1.5 * ps && bs < 2 * ps : rate increases to max. rate
3. bs = n * ps ; (n >= 2) : ~1.1GB/s (== max. rate)
4. bs > n * ps && bs < ~(n+0.5) * ps ; (n > 2) : ~27-70MB/s (differs with ps)
5. bs > ~(n+0.5) * ps && bs < (n+1) * ps ; (n > 2) : increasing rate in several, more or
less, distinct steps (e.g. 1/3 of max. rate and then 2/3 of max rate for 64k pages)

I've tested all four possible page sizes on Itanium (4k, 8k, 16k and 64k) and the pattern is always the same!!

With kernel 2.6.9 (any kernel before 2.6.10-bk6) the read rate is always at ~1.1GB/s, independent of the block size.

This simple patch solves the problem, but I have no idea of possible side-effects ...

--- linux-2.6.12-rc2_orig/mm/filemap.c 2005-04-04 18:40:05.000000000 +0200
+++ linux-2.6.12-rc2/mm/filemap.c 2005-04-20 10:27:42.000000000 +0200

mm/filemap.c 2005-04-20 10:27:42.000000000 +0200 renamed from mm/filemap.c 2005-04-04 18:40:05.000000000 +0200 +1 -1
@@ -719,7 +719,7 @@
 index = *ppos >> PAGE_CACHE_SHIFT;
 next_index = index;
 prev_index = ra.prev_page;
- last_index = (*ppos + desc->count + PAGE_CACHE_SIZE-1) >> 
 PAGE_CACHE_SHIFT;
+ last_index = (*ppos + desc->count + PAGE_CACHE_SIZE) >> 
 PAGE_CACHE_SHIFT;
 offset = *ppos & ~PAGE_CACHE_MASK;

isize = i_size_read(inode);
--- linux-2.6.12-rc2_orig/mm/readahead.c 2005-04-04 18:40:05.000000000 +0200
+++ linux-2.6.12-rc2/mm/readahead.c 2005-04-20 18:37:04.000000000 +0200

(quoted diff) +1 -1
@@ -70,7 +70,7 @@
 */
 static unsigned long get_init_ra_size(unsigned long size, unsigned long max)
 {
- unsigned long newsize = roundup_pow_of_two(size);
+ unsigned long newsize = size;
 
 if (newsize <= max / 64)
 newsize = newsize * newsize;

In order to keep this mail short, I've created a webpage that contains all the detailed information and some plots:
http://www.cern.ch/openlab-debugging/raid

Regards,

Andreas Hirstius

-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/

#2 lore raw

Just tried it, but the performance problem remains :-(
(actually, why should it change? This part of the code didn't change so much between 2.6.10-bk6 and -bk7...)

Andreas

Quoted 115 lines · jmerkey

For 3Ware, you need to chage the queue depths, and you will see dramatically improved performance. 3Ware can take requests a lot faster than Linux pushes them out. Try changing this instead, you won't be going to sleep all the time waiting on the read/write request queues to get "unstarved".

/linux/include/linux/blkdev.h

//#define BLKDEV_MIN_RQ 4
//#define BLKDEV_MAX_RQ 128 /* Default maximum */
#define BLKDEV_MIN_RQ 4096
#define BLKDEV_MAX_RQ 8192 /* Default maximum */

Jeff

Quoted 95 lines · Andreas Hirstius

Hi,

We have a rx4640 with 3x 3Ware 9500 SATA controllers and 24x WD740GD HDD in a software RAID0 configuration (using md). With kernel 2.6.11 the read performance on the md is reduced by a factor of 20 (!!) compared to previous kernels. The write rate to the md doesn't change!! (it actually improves a bit).

The config for the kernels are basically identical.

Here is some vmstat output:

kernel 2.6.9: ~1GB/s read
procs memory swap io system cpu
r b swpd free buff cache si so bi bo in cs us sy wa id 1 1 0 12672 6592 15914112 0 0 1081344 56 15719 1583 0 11 14 74 1 0 0 12672 6592 15915200 0 0 1130496 0 15996 1626 0 11 14 74 0 1 0 12672 6592 15914112 0 0 1081344 0 15891 1570 0 11 14 74 0 1 0 12480 6592 15914112 0 0 1081344 0 15855 1537 0 11 14 74 1 0 0 12416 6592 15914112 0 0 1130496 0 16006 1586 0 12 14 74

kernel 2.6.11: ~55MB/s read
procs memory swap io system cpu
r b swpd free buff cache si so bi bo in cs us sy wa id 1 1 0 24448 37568 15905984 0 0 56934 0 5166 1862 0 1 24 75 0 1 0 20672 37568 15909248 0 0 57280 0 5168 1871 0 1 24 75 0 1 0 22848 37568 15907072 0 0 57306 0 5173 1874 0 1 24 75 0 1 0 25664 37568 15903808 0 0 57190 0 5171 1870 0 1 24 75 0 1 0 21952 37568 15908160 0 0 57267 0 5168 1871 0 1 24 75

Because the filesystem might have an impact on the measurement, "dd" on /dev/md0
was used to get information about the performance. This also opens the possibility to test with block sizes larger than the page size. And it appears that the performance with kernel 2.6.11 is closely related to the block size.
For example if the block size is exactly a multiple (>2) of the page size the performance is back to ~1.1GB/s. The general behaviour is a bit more complicated:
1. bs <= 1.5 * ps : ~27-57MB/s (differs with ps)
2. bs > 1.5 * ps && bs < 2 * ps : rate increases to max. rate
3. bs = n * ps ; (n >= 2) : ~1.1GB/s (== max. rate)
4. bs > n * ps && bs < ~(n+0.5) * ps ; (n > 2) : ~27-70MB/s (differs with ps)
5. bs > ~(n+0.5) * ps && bs < (n+1) * ps ; (n > 2) : increasing rate in several, more or
less, distinct steps (e.g. 1/3 of max. rate and then 2/3 of max rate for 64k pages)

I've tested all four possible page sizes on Itanium (4k, 8k, 16k and 64k) and the pattern is always the same!!

With kernel 2.6.9 (any kernel before 2.6.10-bk6) the read rate is always at ~1.1GB/s, independent of the block size.

This simple patch solves the problem, but I have no idea of possible side-effects ...

--- linux-2.6.12-rc2_orig/mm/filemap.c 2005-04-04 18:40:05.000000000 +0200
+++ linux-2.6.12-rc2/mm/filemap.c 2005-04-20 10:27:42.000000000 +0200

mm/filemap.c 2005-04-20 10:27:42.000000000 +0200 renamed from mm/filemap.c 2005-04-04 18:40:05.000000000 +0200 +1 -1
@@ -719,7 +719,7 @@
 index = *ppos >> PAGE_CACHE_SHIFT;
 next_index = index;
 prev_index = ra.prev_page;
- last_index = (*ppos + desc->count + PAGE_CACHE_SIZE-1) >> 
 PAGE_CACHE_SHIFT;
+ last_index = (*ppos + desc->count + PAGE_CACHE_SIZE) >> 
 PAGE_CACHE_SHIFT;
 offset = *ppos & ~PAGE_CACHE_MASK;

isize = i_size_read(inode);
--- linux-2.6.12-rc2_orig/mm/readahead.c 2005-04-04 18:40:05.000000000 +0200
+++ linux-2.6.12-rc2/mm/readahead.c 2005-04-20 18:37:04.000000000 +0200

(quoted diff) +1 -1
@@ -70,7 +70,7 @@
 */
 static unsigned long get_init_ra_size(unsigned long size, unsigned long max)
 {
- unsigned long newsize = roundup_pow_of_two(size);
+ unsigned long newsize = size;
 
 if (newsize <= max / 64)
 newsize = newsize * newsize;

In order to keep this mail short, I've created a webpage that contains all the detailed information and some plots:
http://www.cern.ch/openlab-debugging/raid

Regards,

Andreas Hirstius

-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/

#3 lore raw
In reply to jmerkey

I was curious if your patch would change the write rate because I see only ~550MB/s (continuous) which is about a factor two away from the capabilities of the disks. ... and got this behaviour (with and without my other patch):

(with single "dd if=/dev/zero of=testxx bs=65536 count=150000 &" or several of them in parallel on an XFS fs)

"vmstat 1" output

0  0      0  28416  37888 15778368    0    0     0     0 8485  3043  0  0  0 100
6  0      0  22144  37952 15785920    0    0     0 12356 7695  2029  0 61  0 39
7  0      0  20864  38016 15785856    0    0   324 1722240 8046  4159  0 100  0  0
7  0      0  20864  38016 15784768    0    0     0 1261440 8391  5222  0 100  0  0
7  0      0  25984  38016 15781504    0    0     0 2003456 8372  5038  0 100  0  0
0  6      0  22784  38016 15781504    0    0     0 2826624 8397  8423  0 93  7  0
0  0      0  21632  38016 15783680    0    0     0 147840 8572 12114  0  9 17 74
0  0      0  21632  38016 15783680    0    0     0    52 8586  5185  0  0  0 100
0  0      0  21632  38016 15783680    0    0     0     0 8588  5412  0  0  0 100
0  0      0  21632  38016 15783680    0    0     0     0 8580  5372  0  0  0 100
0  0      0  21632  38016 15783680    0    0     0     0 7840  5590  0  0  0 100
0  0      0  21632  38016 15783680    0    0     0     0 8587  5321  0  0  0 100
0  0      0  21632  38016 15783680    0    0     0     0 8569  5575  0  0  0 100
0  0      0  21632  38016 15783680    0    0     0     0 8550  5157  0  0  0 100
0  0      0  21632  38016 15783680    0    0     0     0 7963  5640  0  0  0 100
0  0      0  21632  38016 15783680    0    0     0    32 8583  4434  0  0  0 100
7  0      0  20800  38016 15784768    0    0     0  7424 8404  3638  0 15  0 85
8  0      0  20864  38016 15786944    0    0     0 688768 7357  3221  0 100  0  0
8  0      0  20736  28544 15794240    0    0     0 1978560 8376  4897  0 100  0  0
7  0      0  22208  20736 15798784    0    0     0 1385088 8367  4984  0 100  0  0
6  0      0  22144   6848 15812672    0    0    56 1291904 8377  4815  0 100  0  0
0  0      0  50240   6848 15809408    0    0   304  3136 8556  5088  1 26  0 74
0  0      0  50304   6848 15809408    0    0     0     0 8572  5181  0  0  0 100

The average rate here is again pretty close to 550MB/s, it just writes the blocks in "bursts"...

Andreas

Quoted 115 lines · jmerkey

For 3Ware, you need to chage the queue depths, and you will see dramatically improved performance. 3Ware can take requests a lot faster than Linux pushes them out. Try changing this instead, you won't be going to sleep all the time waiting on the read/write request queues to get "unstarved".

/linux/include/linux/blkdev.h

//#define BLKDEV_MIN_RQ 4
//#define BLKDEV_MAX_RQ 128 /* Default maximum */
#define BLKDEV_MIN_RQ 4096
#define BLKDEV_MAX_RQ 8192 /* Default maximum */

Jeff

Quoted 95 lines · Andreas Hirstius

Hi,

We have a rx4640 with 3x 3Ware 9500 SATA controllers and 24x WD740GD HDD in a software RAID0 configuration (using md). With kernel 2.6.11 the read performance on the md is reduced by a factor of 20 (!!) compared to previous kernels. The write rate to the md doesn't change!! (it actually improves a bit).

The config for the kernels are basically identical.

Here is some vmstat output:

kernel 2.6.9: ~1GB/s read
procs memory swap io system cpu
r b swpd free buff cache si so bi bo in cs us sy wa id 1 1 0 12672 6592 15914112 0 0 1081344 56 15719 1583 0 11 14 74 1 0 0 12672 6592 15915200 0 0 1130496 0 15996 1626 0 11 14 74 0 1 0 12672 6592 15914112 0 0 1081344 0 15891 1570 0 11 14 74 0 1 0 12480 6592 15914112 0 0 1081344 0 15855 1537 0 11 14 74 1 0 0 12416 6592 15914112 0 0 1130496 0 16006 1586 0 12 14 74

kernel 2.6.11: ~55MB/s read
procs memory swap io system cpu
r b swpd free buff cache si so bi bo in cs us sy wa id 1 1 0 24448 37568 15905984 0 0 56934 0 5166 1862 0 1 24 75 0 1 0 20672 37568 15909248 0 0 57280 0 5168 1871 0 1 24 75 0 1 0 22848 37568 15907072 0 0 57306 0 5173 1874 0 1 24 75 0 1 0 25664 37568 15903808 0 0 57190 0 5171 1870 0 1 24 75 0 1 0 21952 37568 15908160 0 0 57267 0 5168 1871 0 1 24 75

Because the filesystem might have an impact on the measurement, "dd" on /dev/md0
was used to get information about the performance. This also opens the possibility to test with block sizes larger than the page size. And it appears that the performance with kernel 2.6.11 is closely related to the block size.
For example if the block size is exactly a multiple (>2) of the page size the performance is back to ~1.1GB/s. The general behaviour is a bit more complicated:
1. bs <= 1.5 * ps : ~27-57MB/s (differs with ps)
2. bs > 1.5 * ps && bs < 2 * ps : rate increases to max. rate
3. bs = n * ps ; (n >= 2) : ~1.1GB/s (== max. rate)
4. bs > n * ps && bs < ~(n+0.5) * ps ; (n > 2) : ~27-70MB/s (differs with ps)
5. bs > ~(n+0.5) * ps && bs < (n+1) * ps ; (n > 2) : increasing rate in several, more or
less, distinct steps (e.g. 1/3 of max. rate and then 2/3 of max rate for 64k pages)

I've tested all four possible page sizes on Itanium (4k, 8k, 16k and 64k) and the pattern is always the same!!

With kernel 2.6.9 (any kernel before 2.6.10-bk6) the read rate is always at ~1.1GB/s, independent of the block size.

This simple patch solves the problem, but I have no idea of possible side-effects ...

--- linux-2.6.12-rc2_orig/mm/filemap.c 2005-04-04 18:40:05.000000000 +0200
+++ linux-2.6.12-rc2/mm/filemap.c 2005-04-20 10:27:42.000000000 +0200

mm/filemap.c 2005-04-20 10:27:42.000000000 +0200 renamed from mm/filemap.c 2005-04-04 18:40:05.000000000 +0200 +1 -1
@@ -719,7 +719,7 @@
 index = *ppos >> PAGE_CACHE_SHIFT;
 next_index = index;
 prev_index = ra.prev_page;
- last_index = (*ppos + desc->count + PAGE_CACHE_SIZE-1) >> 
 PAGE_CACHE_SHIFT;
+ last_index = (*ppos + desc->count + PAGE_CACHE_SIZE) >> 
 PAGE_CACHE_SHIFT;
 offset = *ppos & ~PAGE_CACHE_MASK;

isize = i_size_read(inode);
--- linux-2.6.12-rc2_orig/mm/readahead.c 2005-04-04 18:40:05.000000000 +0200
+++ linux-2.6.12-rc2/mm/readahead.c 2005-04-20 18:37:04.000000000 +0200

(quoted diff) +1 -1
@@ -70,7 +70,7 @@
 */
 static unsigned long get_init_ra_size(unsigned long size, unsigned long max)
 {
- unsigned long newsize = roundup_pow_of_two(size);
+ unsigned long newsize = size;
 
 if (newsize <= max / 64)
 newsize = newsize * newsize;

In order to keep this mail short, I've created a webpage that contains all the detailed information and some plots:
http://www.cern.ch/openlab-debugging/raid

Regards,

Andreas Hirstius

-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/

#4 lore raw

Burst is good. There's another window in the SCSI layer that limits to bursts of 128 sector runs (this seems to be the behavior on 3Ware). I've never changed this, but increasing the max number of SCSI requests at this layer may help. The bursty behavior is good, BTW.

Jeff

Quoted 160 lines · Andreas Hirstius

I was curious if your patch would change the write rate because I see only ~550MB/s (continuous) which is about a factor two away from the capabilities of the disks. ... and got this behaviour (with and without my other patch):

(with single "dd if=/dev/zero of=testxx bs=65536 count=150000 &" or several of them in parallel on an XFS fs)

"vmstat 1" output

0  0      0  28416  37888 15778368    0    0     0     0 8485  3043  0  0  0 100
6  0      0  22144  37952 15785920    0    0     0 12356 7695  2029  0 61  0 39
7  0      0  20864  38016 15785856    0    0   324 1722240 8046  4159  0 100  0  0
7  0      0  20864  38016 15784768    0    0     0 1261440 8391  5222  0 100  0  0
7  0      0  25984  38016 15781504    0    0     0 2003456 8372  5038  0 100  0  0
0  6      0  22784  38016 15781504    0    0     0 2826624 8397  8423  0 93  7  0
0  0      0  21632  38016 15783680    0    0     0 147840 8572 12114  0  9 17 74
0  0      0  21632  38016 15783680    0    0     0    52 8586  5185  0  0  0 100
0  0      0  21632  38016 15783680    0    0     0     0 8588  5412  0  0  0 100
0  0      0  21632  38016 15783680    0    0     0     0 8580  5372  0  0  0 100
0  0      0  21632  38016 15783680    0    0     0     0 7840  5590  0  0  0 100
0  0      0  21632  38016 15783680    0    0     0     0 8587  5321  0  0  0 100
0  0      0  21632  38016 15783680    0    0     0     0 8569  5575  0  0  0 100
0  0      0  21632  38016 15783680    0    0     0     0 8550  5157  0  0  0 100
0  0      0  21632  38016 15783680    0    0     0     0 7963  5640  0  0  0 100
0  0      0  21632  38016 15783680    0    0     0    32 8583  4434  0  0  0 100
7  0      0  20800  38016 15784768    0    0     0  7424 8404  3638  0 15  0 85
8  0      0  20864  38016 15786944    0    0     0 688768 7357  3221  0 100  0  0
8  0      0  20736  28544 15794240    0    0     0 1978560 8376  4897  0 100  0  0
7  0      0  22208  20736 15798784    0    0     0 1385088 8367  4984  0 100  0  0
6  0      0  22144   6848 15812672    0    0    56 1291904 8377  4815  0 100  0  0
0  0      0  50240   6848 15809408    0    0   304  3136 8556  5088  1 26  0 74
0  0      0  50304   6848 15809408    0    0     0     0 8572  5181  0  0  0 100

The average rate here is again pretty close to 550MB/s, it just writes the blocks in "bursts"...

Andreas

Quoted 108 lines · jmerkey

For 3Ware, you need to chage the queue depths, and you will see dramatically improved performance. 3Ware can take requests a lot faster than Linux pushes them out. Try changing this instead, you won't be going to sleep all the time waiting on the read/write request queues to get "unstarved".

/linux/include/linux/blkdev.h

//#define BLKDEV_MIN_RQ 4
//#define BLKDEV_MAX_RQ 128 /* Default maximum */
#define BLKDEV_MIN_RQ 4096
#define BLKDEV_MAX_RQ 8192 /* Default maximum */

Jeff

Quoted 89 lines · Andreas Hirstius

Hi,

We have a rx4640 with 3x 3Ware 9500 SATA controllers and 24x WD740GD HDD in a software RAID0 configuration (using md). With kernel 2.6.11 the read performance on the md is reduced by a factor of 20 (!!) compared to previous kernels. The write rate to the md doesn't change!! (it actually improves a bit).

The config for the kernels are basically identical.

Here is some vmstat output:

kernel 2.6.9: ~1GB/s read
procs memory swap io system cpu
r b swpd free buff cache si so bi bo in cs us sy wa id 1 1 0 12672 6592 15914112 0 0 1081344 56 15719 1583 0 11 14 74 1 0 0 12672 6592 15915200 0 0 1130496 0 15996 1626 0 11 14 74 0 1 0 12672 6592 15914112 0 0 1081344 0 15891 1570 0 11 14 74 0 1 0 12480 6592 15914112 0 0 1081344 0 15855 1537 0 11 14 74 1 0 0 12416 6592 15914112 0 0 1130496 0 16006 1586 0 12 14 74

kernel 2.6.11: ~55MB/s read
procs memory swap io system cpu
r b swpd free buff cache si so bi bo in cs us sy wa id 1 1 0 24448 37568 15905984 0 0 56934 0 5166 1862 0 1 24 75 0 1 0 20672 37568 15909248 0 0 57280 0 5168 1871 0 1 24 75 0 1 0 22848 37568 15907072 0 0 57306 0 5173 1874 0 1 24 75 0 1 0 25664 37568 15903808 0 0 57190 0 5171 1870 0 1 24 75 0 1 0 21952 37568 15908160 0 0 57267 0 5168 1871 0 1 24 75

Because the filesystem might have an impact on the measurement, "dd" on /dev/md0
was used to get information about the performance. This also opens the possibility to test with block sizes larger than the page size. And it appears that the performance with kernel 2.6.11 is closely related to the block size.
For example if the block size is exactly a multiple (>2) of the page size the performance is back to ~1.1GB/s. The general behaviour is a bit more complicated:
1. bs <= 1.5 * ps : ~27-57MB/s (differs with ps)
2. bs > 1.5 * ps && bs < 2 * ps : rate increases to max. rate
3. bs = n * ps ; (n >= 2) : ~1.1GB/s (== max. rate)
4. bs > n * ps && bs < ~(n+0.5) * ps ; (n > 2) : ~27-70MB/s (differs with ps)
5. bs > ~(n+0.5) * ps && bs < (n+1) * ps ; (n > 2) : increasing rate in several, more or
less, distinct steps (e.g. 1/3 of max. rate and then 2/3 of max rate for 64k pages)

I've tested all four possible page sizes on Itanium (4k, 8k, 16k and 64k) and the pattern is always the same!!

With kernel 2.6.9 (any kernel before 2.6.10-bk6) the read rate is always at ~1.1GB/s, independent of the block size.

This simple patch solves the problem, but I have no idea of possible side-effects ...

--- linux-2.6.12-rc2_orig/mm/filemap.c 2005-04-04 18:40:05.000000000 +0200
+++ linux-2.6.12-rc2/mm/filemap.c 2005-04-20 10:27:42.000000000 +0200

mm/filemap.c 2005-04-20 10:27:42.000000000 +0200 renamed from mm/filemap.c 2005-04-04 18:40:05.000000000 +0200 +1 -1
@@ -719,7 +719,7 @@
 index = *ppos >> PAGE_CACHE_SHIFT;
 next_index = index;
 prev_index = ra.prev_page;
- last_index = (*ppos + desc->count + PAGE_CACHE_SIZE-1) >> 
 PAGE_CACHE_SHIFT;
+ last_index = (*ppos + desc->count + PAGE_CACHE_SIZE) >> 
 PAGE_CACHE_SHIFT;
 offset = *ppos & ~PAGE_CACHE_MASK;

isize = i_size_read(inode);
--- linux-2.6.12-rc2_orig/mm/readahead.c 2005-04-04 18:40:05.000000000 +0200
+++ linux-2.6.12-rc2/mm/readahead.c 2005-04-20 18:37:04.000000000 +0200

(quoted diff) +1 -1
@@ -70,7 +70,7 @@
 */
 static unsigned long get_init_ra_size(unsigned long size, unsigned long max)
 {
- unsigned long newsize = roundup_pow_of_two(size);
+ unsigned long newsize = size;
 
 if (newsize <= max / 64)
 newsize = newsize * newsize;

In order to keep this mail short, I've created a webpage that contains all the detailed information and some plots:
http://www.cern.ch/openlab-debugging/raid

Regards,

Andreas Hirstius

Quoted 7 lines
Quoted 6 lines

-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/

-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/

#5 lore raw
In reply to jmerkey
Quoted 8 lines · jmerkey

For 3Ware, you need to chage the queue depths, and you will see dramatically improved performance. 3Ware can take requests
a lot faster than Linux pushes them out. Try changing this instead, you won't be going to sleep all the time waiting on the read/write request queues to get "unstarved".

Quoted 7 lines

/linux/include/linux/blkdev.h

//#define BLKDEV_MIN_RQ 4
//#define BLKDEV_MAX_RQ 128 /* Default maximum */
#define BLKDEV_MIN_RQ 4096
#define BLKDEV_MAX_RQ 8192 /* Default maximum */

BTW, don't do this. BLKDEV_MIN_RQ sets the size of the mempool reserved requests and will only get slightly used in low memory conditions, so most memory will probably be wasted.

Just change /sys/block/xxx/queue/nr_requests

Nick

#6 lore raw
In reply to jmerkey

A small update.

Patching mm/filemap.c is not necessary in order to get the improved performance! It's sufficient to remove roundup_pow_of_two from |get_init_ra_size ...

So a simple one-liner changes to picture dramatically. But why ?!?!?

Andreas
|

Quoted 115 lines · jmerkey

For 3Ware, you need to chage the queue depths, and you will see dramatically improved performance. 3Ware can take requests a lot faster than Linux pushes them out. Try changing this instead, you won't be going to sleep all the time waiting on the read/write request queues to get "unstarved".

/linux/include/linux/blkdev.h

//#define BLKDEV_MIN_RQ 4
//#define BLKDEV_MAX_RQ 128 /* Default maximum */
#define BLKDEV_MIN_RQ 4096
#define BLKDEV_MAX_RQ 8192 /* Default maximum */

Jeff

Quoted 95 lines · Andreas Hirstius

Hi,

We have a rx4640 with 3x 3Ware 9500 SATA controllers and 24x WD740GD HDD in a software RAID0 configuration (using md). With kernel 2.6.11 the read performance on the md is reduced by a factor of 20 (!!) compared to previous kernels. The write rate to the md doesn't change!! (it actually improves a bit).

The config for the kernels are basically identical.

Here is some vmstat output:

kernel 2.6.9: ~1GB/s read
procs memory swap io system cpu
r b swpd free buff cache si so bi bo in cs us sy wa id 1 1 0 12672 6592 15914112 0 0 1081344 56 15719 1583 0 11 14 74 1 0 0 12672 6592 15915200 0 0 1130496 0 15996 1626 0 11 14 74 0 1 0 12672 6592 15914112 0 0 1081344 0 15891 1570 0 11 14 74 0 1 0 12480 6592 15914112 0 0 1081344 0 15855 1537 0 11 14 74 1 0 0 12416 6592 15914112 0 0 1130496 0 16006 1586 0 12 14 74

kernel 2.6.11: ~55MB/s read
procs memory swap io system cpu
r b swpd free buff cache si so bi bo in cs us sy wa id 1 1 0 24448 37568 15905984 0 0 56934 0 5166 1862 0 1 24 75 0 1 0 20672 37568 15909248 0 0 57280 0 5168 1871 0 1 24 75 0 1 0 22848 37568 15907072 0 0 57306 0 5173 1874 0 1 24 75 0 1 0 25664 37568 15903808 0 0 57190 0 5171 1870 0 1 24 75 0 1 0 21952 37568 15908160 0 0 57267 0 5168 1871 0 1 24 75

Because the filesystem might have an impact on the measurement, "dd" on /dev/md0
was used to get information about the performance. This also opens the possibility to test with block sizes larger than the page size. And it appears that the performance with kernel 2.6.11 is closely related to the block size.
For example if the block size is exactly a multiple (>2) of the page size the performance is back to ~1.1GB/s. The general behaviour is a bit more complicated:
1. bs <= 1.5 * ps : ~27-57MB/s (differs with ps)
2. bs > 1.5 * ps && bs < 2 * ps : rate increases to max. rate
3. bs = n * ps ; (n >= 2) : ~1.1GB/s (== max. rate)
4. bs > n * ps && bs < ~(n+0.5) * ps ; (n > 2) : ~27-70MB/s (differs with ps)
5. bs > ~(n+0.5) * ps && bs < (n+1) * ps ; (n > 2) : increasing rate in several, more or
less, distinct steps (e.g. 1/3 of max. rate and then 2/3 of max rate for 64k pages)

I've tested all four possible page sizes on Itanium (4k, 8k, 16k and 64k) and the pattern is always the same!!

With kernel 2.6.9 (any kernel before 2.6.10-bk6) the read rate is always at ~1.1GB/s, independent of the block size.

This simple patch solves the problem, but I have no idea of possible side-effects ...

--- linux-2.6.12-rc2_orig/mm/filemap.c 2005-04-04 18:40:05.000000000 +0200
+++ linux-2.6.12-rc2/mm/filemap.c 2005-04-20 10:27:42.000000000 +0200

mm/filemap.c 2005-04-20 10:27:42.000000000 +0200 renamed from mm/filemap.c 2005-04-04 18:40:05.000000000 +0200 +1 -1
@@ -719,7 +719,7 @@
 index = *ppos >> PAGE_CACHE_SHIFT;
 next_index = index;
 prev_index = ra.prev_page;
- last_index = (*ppos + desc->count + PAGE_CACHE_SIZE-1) >> 
 PAGE_CACHE_SHIFT;
+ last_index = (*ppos + desc->count + PAGE_CACHE_SIZE) >> 
 PAGE_CACHE_SHIFT;
 offset = *ppos & ~PAGE_CACHE_MASK;

isize = i_size_read(inode);
--- linux-2.6.12-rc2_orig/mm/readahead.c 2005-04-04 18:40:05.000000000 +0200
+++ linux-2.6.12-rc2/mm/readahead.c 2005-04-20 18:37:04.000000000 +0200

(quoted diff) +1 -1
@@ -70,7 +70,7 @@
 */
 static unsigned long get_init_ra_size(unsigned long size, unsigned long max)
 {
- unsigned long newsize = roundup_pow_of_two(size);
+ unsigned long newsize = size;
 
 if (newsize <= max / 64)
 newsize = newsize * newsize;

In order to keep this mail short, I've created a webpage that contains all the detailed information and some plots:
http://www.cern.ch/openlab-debugging/raid

Regards,

Andreas Hirstius

-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/

#8 lore raw
In reply to a message not in the archive

Hi!

Quoted 8 lines

A small update.

Patching mm/filemap.c is not necessary in order to get the improved performance!
It's sufficient to remove roundup_pow_of_two from |get_init_ra_size ...

So a simple one-liner changes to picture dramatically. But why ?!?!?

roundup_pow_of_two() uses fls() and ia64 has buggy fls() implementation [ seems that David fixed it but patch is not in the mainline yet]:

http://www.mail-archive.com/[email protected]/msg01196.html

That would also explain why you couldn't reproduce the problem on ia32 Xeon machines.

Bartlomiej

#9 lore raw

Hi,

The fls() patch from David solves the problem :-))

Do you have an idea, when it will be in the mainline kernel??

Andreas

Quoted 22 lines · Bartlomiej ZOLNIERKIEWICZ

Hi!

Quoted 8 lines

A small update.

Patching mm/filemap.c is not necessary in order to get the improved performance!
It's sufficient to remove roundup_pow_of_two from |get_init_ra_size ...

So a simple one-liner changes to picture dramatically. But why ?!?!?

roundup_pow_of_two() uses fls() and ia64 has buggy fls() implementation [ seems that David fixed it but patch is not in the mainline yet]:

http://www.mail-archive.com/[email protected]/msg01196.html

That would also explain why you couldn't reproduce the problem on ia32 Xeon machines.

Bartlomiej

#10 lore raw

Tony and Andrew,

I just checked 2.6.12-rc3 and the fls() fix is indeed missing. Do you know what happened?

	--david
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line · Andreas Hirstius

Andreas> Hi, The fls() patch from David solves the problem :-))

Andreas> Do you have an idea, when it will be in the mainline Andreas> kernel??

Andreas> Andreas

Quoted 16 lines · Andreas> Bartlomiej ZOLNIERKIEWICZ wrote:
Quoted 16 lines

Hi!

Quoted 8 lines

A small update.

Patching mm/filemap.c is not necessary in order to get the improved performance! It's sufficient to remove roundup_pow_of_two from |get_init_ra_size ...

So a simple one-liner changes to picture dramatically. But why ?!?!?

roundup_pow_of_two() uses fls() and ia64 has buggy fls() implementation [ seems that David fixed it but patch is not in the mainline yet]:

Quoted 7 lines
Quoted 7 lines

http://www.mail-archive.com/[email protected]/msg01196.html

That would also explain why you couldn't reproduce the problem on ia32 Xeon machines.

Bartlomiej

Andreas> _______________________________________________ Andreas> Gelato-technical mailing list
Andreas> [email protected]
Andreas> https://www.gelato.unsw.edu.au/mailman/listinfo/gelato-technical