Re: [Bug #13463] Poor SSD performance

From: Jake Ellowitz
Date: Tue Jun 16 2009 - 00:30:12 EST

Next message: Davide Libenzi: "Re: 2.6.30-rc1 regression? -- epoll: BUG: sleeping function calledfrom invalid context"
Previous message: starlight: "Re: QUESTION: can netdev_alloc_skb() errors be reduced by tuning?"
Next in thread: Wu Fengguang: "Re: [Bug #13463] Poor SSD performance"
Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]

Dear Fengguang,

Thanks so much for the attention you paid to this problem. I did not want to respond until I got a chance to give the new kernel a shot to see if the bug was still present. It appears not to be -- hdparm and dd both register read speeds between 200 and 220 MB/s as opposed to the 70 to 80 MB/s I was getting with kernel 2.6.29. So, I guess this strange bug has sort of resolved itself.

Best,
Jake

Wu Fengguang wrote:

On Wed, Jun 10, 2009 at 02:37:46PM +0800, Wu Fengguang wrote:

Bug-Entry : http://bugzilla.kernel.org/show_bug.cgi?id=13463
Subject : Poor SSD performance
Submitter : Jake <ellowitz@xxxxxxxxxxxx>
Date : 2009-06-05 17:37 (3 days old)

Hi Jake,

Could you collect some blktrace data for the dd commands on new/old
kernels?

dd if=/dev/sda of=/dev/null bs=1M count=1024 iflag=direct
dd if=/dev/sda of=/dev/null bs=1M count=1024

I managed to get a SanDisk SSD for testing, and observes that

- one must increase read_ahead_kb to at least max_sectors_kb or better
"bs=1M" to make a fair comparison
- with increased readahead size, the dd reported throughputs are
75MB/s vs 77MB/s, while the blktrace reported throughputs are
75MB/s vs 75MB/s (buffered IO vs direct IO).

Here are details.

The dd throughputs are equal for rotational hard disks, but differs
for this SanDisk SSD (with default RA parameters):

% dd if=/dev/sda of=/dev/null iflag=direct bs=1M count=1024
1073741824 bytes (1.1 GB) copied, 13.905 s, 77.2 MB/s
1073741824 bytes (1.1 GB) copied, 13.9029 s, 77.2 MB/s

% dd if=/dev/sda of=/dev/null bs=1M count=1024 1073741824 bytes (1.1 GB) copied, 14.7294 s, 72.9 MB/s
1073741824 bytes (1.1 GB) copied, 14.8647 s, 72.2 MB/s

Here is the blktrace summary:

dd dd-direct
--------------------------------------------------------------------------------------
CPU0 (sda): | CPU0 (sda):
Reads Queued: 9,888, 39,552KiB | Reads Queued: 84, 43,008KiB Read Dispatches: 302, 38,588KiB | Read Dispatches: 84, 43,008KiB Reads Requeued: 0 | Reads Requeued: 0 Reads Completed: 337, 44,600KiB | Reads Completed: 83, 42,496KiB Read Merges: 9,574, 38,296KiB | Read Merges: 0, 0KiB Read depth: 2 | Read depth: 2 IO unplugs: 313 | IO unplugs: 42 CPU1 (sda): | CPU1 (sda):
Reads Queued: 11,840, 47,360KiB | Reads Queued: 96, 49,152KiB Read Dispatches: 372, 48,196KiB | Read Dispatches: 96, 49,152KiB Reads Requeued: 0 | Reads Requeued: 0 Reads Completed: 337, 42,312KiB | Reads Completed: 96, 49,152KiB Read Merges: 11,479, 45,916KiB | Read Merges: 0, 0KiB Read depth: 2 | Read depth: 2 IO unplugs: 372 | IO unplugs: 48 | Total (sda): | Total (sda):
Reads Queued: 21,728, 86,912KiB | Reads Queued: 180, 92,160KiB Read Dispatches: 674, 86,784KiB | Read Dispatches: 180, 92,160KiB Reads Requeued: 0 | Reads Requeued: 0 Reads Completed: 674, 86,912KiB | Reads Completed: 179, 91,648KiB Read Merges: 21,053, 84,212KiB | Read Merges: 0, 0KiB IO unplugs: 685 | IO unplugs: 90 | Throughput (R/W): 69,977KiB/s / 0KiB/s | Throughput (R/W): 75,368KiB/s / 0KiB/s Events (sda): 46,804 entries | Events (sda): 1,158 entries

Another obvious difference is IO size.
One is read_ahead_kb=128K, another is max_sectors_kb=512K:

dd:
8,0 0 13497 0.804939305 0 C R 782592 + 256 [0]
8,0 0 13498 0.806713692 0 C R 782848 + 256 [0]
8,0 1 16275 0.808488708 0 C R 783104 + 256 [0]
8,0 0 13567 0.810261350 0 C R 783360 + 256 [0]
8,0 0 13636 0.812036226 0 C R 783616 + 256 [0]
8,0 1 16344 0.813806353 0 C R 783872 + 256 [0]
8,0 1 16413 0.815578436 0 C R 784128 + 256 [0]
8,0 0 13705 0.817347935 0 C R 784384 + 256 [0]

dd-direct:
8,0 0 428 0.998831975 0 C R 357376 + 1024 [0]
8,0 1 514 1.005683404 0 C R 358400 + 1024 [0]
8,0 1 515 1.012402554 0 C R 359424 + 1024 [0]
8,0 0 440 1.019303850 0 C R 360448 + 1024 [0]
8,0 1 526 1.026024048 0 C R 361472 + 1024 [0]
8,0 1 538 1.032875967 0 C R 362496 + 1024 [0]
8,0 0 441 1.039595815 0 C R 363520 + 1024 [0]

The non-direct dd throughput can improve with 512K and 1M readahead size,
but still a bit slower than the direct dd case:
1073741824 bytes (1.1 GB) copied, 14.1619 s, 75.8 MB/s
1073741824 bytes (1.1 GB) copied, 14.1517 s, 75.9 MB/s

dd-512k dd-direct2
-------------------------------------------------------------------------------------
Total (sda): | Total (sda):
Reads Queued: 23,808, 95,232KiB | Reads Queued: 178, 91,136KiB Read Dispatches: 215, 95,232KiB | Read Dispatches: 178, 91,136KiB Reads Requeued: 0 | Reads Requeued: 0 Reads Completed: 215, 95,232KiB | Reads Completed: 177, 90,624KiB Read Merges: 23,593, 94,372KiB | Read Merges: 0, 0KiB IO unplugs: 236 | IO unplugs: 89 | Throughput (R/W): 75,222KiB/s / 0KiB/s | Throughput (R/W): 75,520KiB/s / 0KiB/s Events (sda): 48,687 entries | Events (sda): 1,145 entries

Interestingly, the throughput reported by blktrace is almost the same,
whereas the dd report favors the dd-direct case.

More parameters.

[ 10.137350] scsi 3:0:0:0: Direct-Access ATA SanDisk SSD SATA 1.13 PQ: 0 ANSI: 5
[ 10.147137] sd 3:0:0:0: [sda] 61500000 512-byte hardware sectors: (31.4 GB/29.3 GiB)
[ 10.155060] sd 3:0:0:0: [sda] Write Protect is off
[ 10.159922] sd 3:0:0:0: [sda] Mode Sense: 00 3a 00 00
[ 10.165179] sd 3:0:0:0: [sda] Write cache: disabled, read cache: enabled, doesn't support DPO or FUA
[ 10.174994] sda:

/dev/sda:

Model=SanDisk SSD SATA 5000 2.5 , FwRev=1.13 , SerialNo= 81402200246
Config={ Fixed }
RawCHS=16383/16/63, TrkSize=0, SectSize=0, ECCbytes=4
BuffType=unknown, BuffSize=0kB, MaxMultSect=1, MultSect=?1?
CurCHS=16383/16/63, CurSects=16514064, LBA=yes, LBAsects=61500000
IORDY=yes, tPIO={min:120,w/IORDY:120}, tDMA={min:120,rec:120}
PIO modes: pio0 pio1 pio2 pio3 pio4 DMA modes: mdma0 mdma1 mdma2 UDMA modes: udma0 udma1 udma2 udma3 udma4 *udma5 AdvancedPM=yes: disabled (255) WriteCache=disabled
Drive conforms to: unknown: ATA/ATAPI-2,3,4,5,6,7

* signifies the current active mode

/sys/block/sda/queue/nr_requests:128
/sys/block/sda/queue/read_ahead_kb:128
/sys/block/sda/queue/max_hw_sectors_kb:32767
/sys/block/sda/queue/max_sectors_kb:512
/sys/block/sda/queue/scheduler:noop [cfq] /sys/block/sda/queue/hw_sector_size:512
/sys/block/sda/queue/rotational:1
/sys/block/sda/queue/nomerges:0
/sys/block/sda/queue/rq_affinity:0
/sys/block/sda/queue/iostats:1
/sys/block/sda/queue/iosched/quantum:4
/sys/block/sda/queue/iosched/fifo_expire_sync:124
/sys/block/sda/queue/iosched/fifo_expire_async:248
/sys/block/sda/queue/iosched/back_seek_max:16384
/sys/block/sda/queue/iosched/back_seek_penalty:2
/sys/block/sda/queue/iosched/slice_sync:100
/sys/block/sda/queue/iosched/slice_async:40
/sys/block/sda/queue/iosched/slice_async_rq:2
/sys/block/sda/queue/iosched/slice_idle:8

Thanks,
Fengguang

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

Next message: Davide Libenzi: "Re: 2.6.30-rc1 regression? -- epoll: BUG: sleeping function calledfrom invalid context"
Previous message: starlight: "Re: QUESTION: can netdev_alloc_skb() errors be reduced by tuning?"
Next in thread: Wu Fengguang: "Re: [Bug #13463] Poor SSD performance"
Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]