Slow perfs with IDE SSD

Slow perfs with IDE SSD

[François-Julien Marbaud]

Hi !

I'm doing an internship in a research lab, and we are trying to use a P.ATA SSD
in order to speed-up a genome computation software.

We got a performance problem, expecially with access time : access time is about
800us, and according to the manufacturer it should be about 40us. The bandwith
is OK.

The SSD is a 2.5' flash hard drive, the slave of the first IDE channel.

We wonder if that could be the because P.ATA driver is not really done for small
access times, so we got such a poor time. In that case, I guess we would have to
have a closer look at the driver ... 

We are also going to do some tests with the SSD on a windows computer, maybe
that's a problem with the firmware of the SSD ...

Thank you for your help !

Here are some details about the configuration.


[rootlav@remix lavenier]# /sbin/hdparm -i /dev/hdb

/dev/hdb:

  Model=M-Systems FFD ULTRA ATA, FwRev=1.13, SerialNo=515240001
  Config={ Fixed }
  RawCHS=16383/16/63, TrkSize=0, SectSize=0, ECCbytes=4
  BuffType=unknown, BuffSize=0kB, MaxMultSect=2, MultSect=off
  CurCHS=16383/16/63, CurSects=16514064, LBA=yes, LBAsects=128817152
  IORDY=yes, tPIO={min:120,w/IORDY:120}, tDMA={min:120,rec:120}
  PIO modes:  pio0 pio1 pio2 pio3 pio4
  DMA modes:  sdma0 sdma1 mdma0 mdma1 mdma2
  UDMA modes: udma0 udma1 udma2 udma3 udma4 *udma5
  AdvancedPM=no WriteCache=enabled
  Drive conforms to: ATA/ATAPI-6 T13 1410D revision 1:

  * signifies the current active mode 

The computer is a DELL precision360 P4 3Ghz
Hard drives : sda SCSI 67MB, hda 80MB, hdb 67MB (the SSD)

OS : Linux Fedora 2 : 2.6.6-1 

The drive is a M-Systems 2,5' FFD ultra-ata flash disk (complies with ATA-6)

http://www.m-systems.com/site/en-US/

3.3. Characteristics
3.3.1. ATA Modes
The Ultra ATA supports the following ATA modes:
• PIO mode 0, 1, 2, 3, 4
• DMA mode 0, 1, 2
• Ultra DMA mode 0, 1, 2, 3, 4, 5
3.3.2. Burst Read/Write Performance
The Ultra ATA burst read/write rate is 100 MB/sec.
3.3.3. Sustained Read/Write Performance
The Ultra ATA meets the performance requirements specified in Table 1.
Performance was measured on a computer using the following setup:
• Capacity: 4GB
• Configuration: Ultra DMA Mode 5, with 128K blocks aligned on 128K boundaries
• Platform: ASUS P4T533-C with Intel 2.4 GHz
• Testing utilities:
o H2Bench (DOS)
o H2benchW (Windows XP)
Table 1: FFD 2.5” Ultra ATA Ultra DMA Mode 5 Transfer Rates
Operation
Burst Rate1 (MB/sec)
Sustained2 Rate 128KBs Blocks (MB/sec)
Read
100.0
45.0
Write
100.0
40.0
3.3.4. Access Time
Maximum access time for the Ultra ATA is < 0.04 msec.
3.3.5. Seek Time
The Ultra ATA has no seek time.
3.3.6. Memory Capacity
Ultra ATA memory capacity information is described in Table 2.



 

Re: Slow perfs with IDE SSD

[Tejun Heo]

François-Julien Marbaud wrote:
> Hi !
> 
> I'm doing an internship in a research lab, and we are trying to use a P.ATA SSD
> in order to speed-up a genome computation software.
> 
> We got a performance problem, expecially with access time : access time is about
> 800us, and according to the manufacturer it should be about 40us. The bandwith
> is OK.
> 
> The SSD is a 2.5' flash hard drive, the slave of the first IDE channel.
> 
> We wonder if that could be the because P.ATA driver is not really done for small
> access times, so we got such a poor time. In that case, I guess we would have to
> have a closer look at the driver ... 

First of all, try noop elevator.

   echo noop > /sys/block/hdx/queue/scheduler

Depending on your workload, adjusting plugging delay/threshold 
parameters might help.  These are initialized in the kernel source file 
block/ll_rw_blk.c::blk_queue_make_request().  q->unplug_thresh and 
q->unplug_delay.

-- 
tejun

Re: Slow perfs with IDE SSD

[Tejun Heo]

Tejun Heo wrote:
> François-Julien Marbaud wrote:
>> Hi !
>>
>> I'm doing an internship in a research lab, and we are trying to use a 
>> P.ATA SSD
>> in order to speed-up a genome computation software.
>>
>> We got a performance problem, expecially with access time : access 
>> time is about
>> 800us, and according to the manufacturer it should be about 40us. The 
>> bandwith
>> is OK.
>>
>> The SSD is a 2.5' flash hard drive, the slave of the first IDE channel.
>>
>> We wonder if that could be the because P.ATA driver is not really done 
>> for small
>> access times, so we got such a poor time. In that case, I guess we 
>> would have to
>> have a closer look at the driver ... 
> 
> First of all, try noop elevator.
> 
>   echo noop > /sys/block/hdx/queue/scheduler
> 
> Depending on your workload, adjusting plugging delay/threshold 
> parameters might help.  These are initialized in the kernel source file 
> block/ll_rw_blk.c::blk_queue_make_request().  q->unplug_thresh and 
> q->unplug_delay.
> 

And one more thing, if a 67MB~4GB SDD drive speeds up the computation, 
something is wrong.  4GB main memory will be much better than 4GB SDD 
for such purposes.

-- 
tejun

Re: Slow perfs with IDE SSD

[François-Julien Marbaud]

Tejun Heo <htejun <at> gmail.com> writes:
> > 
> > First of all, try noop elevator.
> > ...
> 
> And one more thing, if a 67MB~4GB SDD drive speeds up the computation, 
> something is wrong.  4GB main memory will be much better than 4GB SDD 
> for such purposes.
> 



Thank you, we will try that !

(And you are right, that is a 67GB SSD ;)