LSI 53c1030 (Fusion MPT) performance with O_SYNC

LSI 53c1030 (Fusion MPT) performance with O_SYNC

[Jan Oravec]

Hi,


I have got a new dual-Opteron server based on Tyan S2880 motherboard with
LSI53C1030 U320 controller onboard - there are 15kRPM Fujitsu disks.

I've noticed poor performance with MySQL/InnoDB when compared to another
S2880-based box with IDE disks.

I've tracked it to this:

Let's have this code (notice the O_SYNC) and enough large test file.

#include <sys/types.h>
#include <sys/stat.h>
#include <fcntl.h>
#include <unistd.h>

int main(int argc, void **argv)
{
  char buffer[64];
  int fd, i;

  fd=open(argv[1], O_WRONLY|O_SYNC);
  
  lseek(fd, 0, SEEK_SET);
  for(i=0; i<10000; i++)
    write(fd, buffer, 64);
  
  return 0;
}


The results are:

LSI53C1030:
$ time ./a.out testfile 

real    0m4.218s
user    0m0.002s
sys     0m0.219s


IDE:
$ time ./a.out testfile 

real    0m1.767s
user    0m0.002s
sys     0m0.356s


Changing size of written amount of data to 4096 the time ratio between
LSI53C1030 and IDE is around 1:1.


Both boxes are running vanilla 2.6.7 kernel.


Any help appreciated,


Jan


PS: please CC, I am not on the list.

Re: LSI 53c1030 (Fusion MPT) performance with O_SYNC

[Arjan van de Ven]

[-- Attachment #1: Type: text/plain, Size: 336 bytes --]

On Thu, 2004-07-29 at 11:56, Jan Oravec wrote:

> I've noticed poor performance with MySQL/InnoDB when compared to another
> S2880-based box with IDE disks.

your ide disk probably has write back caching enabled while your
mptfusion doesn't..... if you value data integrity over performance the
mptfusion has a saner default ;)

[-- Attachment #2: This is a digitally signed message part --]
[-- Type: application/pgp-signature, Size: 189 bytes --]

Re: LSI 53c1030 (Fusion MPT) performance with O_SYNC

[Jan Oravec]

On Thu, Jul 29, 2004 at 12:04:53PM +0200, Arjan van de Ven wrote:
> On Thu, 2004-07-29 at 11:56, Jan Oravec wrote:
> 
> > I've noticed poor performance with MySQL/InnoDB when compared to another
> > S2880-based box with IDE disks.
> 
> your ide disk probably has write back caching enabled while your
> mptfusion doesn't..... if you value data integrity over performance the
> mptfusion has a saner default ;)

It was enabled:

# sginfo -c /dev/sda
Caching mode page (0x8)
-----------------------
Initiator Control                  0
ABPF                               0
CAP                                0
DISC                               1
SIZE                               0
Write Cache Enabled                1
MF                                 0
Read Cache Disabled                0
Demand Read Retention Priority     0
Demand Write Retention Priority    0
Disable Pre-fetch Transfer Length  65535
Minimum Pre-fetch                  0
Maximum Pre-fetch                  0
Maximum Pre-fetch Ceiling          65535
FSW                                1
LBCSS                              0
DRA                                0
Number of Cache Segments           8
Cache Segment size                 0
Non-Cache Segment size             0

When I've disabled it on both mptfusion and IDE, it took 40s on mpt and 83s
on IDE.

It seems like write-cache is not as effective on mptfusion as it is on IDE?

I am considering plugging there a ZCR with battery, so write-cache has a
sense for me.


Jan