Hi there,

I am trying to understand why my SB1K gives me abysmal performance on
sequential read/write...

My setup is 2x15,000RPM 36GB drives, Solaris 10. I run the Nastran FEA
solver, which sometimes creates huge scratch files that are constantly
accessed... I direct all my scratch files on my second drive... I am
seeing some very long solve time compared to much slower machines.

An example would be SB1K 2x900 being 3 times slower than an sgi Octane2
2x400 for solves that use extensive I/O. For solves with little I/O, the
SB1K is about twice the speed of the Octane2, which is what I expected.
It is not limited to Nastran, other FEA solvers show the same trend...

So, it appears to me that the culprit is the scratch filesystem... I
tried to mount the drive with forcedirectio, but it looks like it made
things even worse... Does anybody have any suggestion?

Somebody on the Sun Hardware forum was saying that this is because the
drives weren't Sun drives, therefore the performance should suck, or at
least wouldn't be as good as it should, with actually no way of making
it any better... Has anybody seen this?

On Fri, 30 Sep 2005 03:46:09 GMT
UNIX Museum <unixmuseum@verizon.com> wrote:

> Somebody on the Sun Hardware forum was saying that this is because
> the drives weren't Sun drives, therefore the performance should suck,
> or at least wouldn't be as good as it should, with actually no way of
> making it any better...

That's baloney, of course.

> Has anybody seen this?

I'm not running the stuff you run, so in order to see how far your disk
throughput is off the mark, I suggest you run a benchmark such as
bonnie on your systems. Here's the output from my SunBlade2000 with
4GB and dual 900MHz CuIII+ CPUs running Solaris9:

-------Sequential Output-------- ---Sequential Input-- --Random--
-Per Char- --Block--- -Rewrite-- -Per Char- --Block--- --Seeks---
MB K/sec %CPU K/sec %CPU K/sec %CPU K/sec %CPU K/sec %CPU /sec %CPU
2000 37452 92.3 46044 34.0 20286 23.7 18077 59.7 26335 15.7 219.3 5.3

A Digital PWS 500au (768MB/500MHz) with SCSI disks running Tru64 5.1A is
far slower:

-------Sequential Output-------- ---Sequential Input-- --Random--
-Per Char- --Block--- -Rewrite-- -Per Char- --Block--- --Seeks---
MB K/sec %CPU K/sec %CPU K/sec %CPU K/sec %CPU K/sec %CPU /sec %CPU
2000 11996 93.8 15477 40.2 4778 11.1 10923 79.1 15849 15.9 129.0 4.5

And a 195MHz Octane with 1GB memory running IRIX 6.5.9 is no match
either:

---Sequential Output-------- ---Sequential Input-- --Random--
-Per Char- --Block--- -Rewrite-- -Per Char- --Block--- --Seeks---
MB K/sec %CPU K/sec %CPU K/sec %CPU K/sec %CPU K/sec %CPU /sec %CPU
2000 8918 95.9 7248 6.7 3558 5.1 7075 93.7 8676 6.2 71.7 1.7

You do have logging switched on, don't you?

Take care,

--
Stefaan
--
As complexity rises, precise statements lose meaning,
and meaningful statements lose precision. -- Lotfi Zadeh

Stefaan A Eeckels wrote:

>
> That's baloney, of course.

Judging how quickly the guy started insulting me on the Sun forum as
soon as I started asking why (felt like 12 year old Linux kids jumping
on a "n00b"), I kind of guessed that... A hard drive is a hard drive,
even if it doesn't have a Sun firmware, this shouldn't be the end of the
world.
>
>
> I'm not running the stuff you run, so in order to see how far your disk
> throughput is off the mark, I suggest you run a benchmark such as
> bonnie on your systems. Here's the output from my SunBlade2000 with
> 4GB and dual 900MHz CuIII+ CPUs running Solaris9:

We're close in configuration, SB1K 2x900 4GB, I'll give bonnie a shot
and will report...

Thanks for your reply, Stefaan.

UNIX Museum wrote:
> Hi there,
>
> I am trying to understand why my SB1K gives me abysmal performance on
> sequential read/write...
>

>
> So, it appears to me that the culprit is the scratch filesystem... I
> tried to mount the drive with forcedirectio, but it looks like it made
> things even worse... Does anybody have any suggestion?

Yepp, forcedirectio is for databases that do their own caching.

If you are doing verylargefiles you should :

1. make the filessystem with fragsize set equal to blocksize
2. set cylinder group size to 32.
3 set the "maxbpg" to 50% or 99% of the amount of blocks
in a cylinder group

Look at fstyp(1M) and tunefs(1M) and newfs(1M)

//Lars

It is interesting to see the difference between the various disk
technologies.
My box is a SB1500 red with 2Gb ram

-------Sequential Output-------- ---Sequential Input--
--Random--
-Per Char- --Block--- -Rewrite-- -Per Char- --Block---
--Seeks---
MB K/sec %CPU K/sec %CPU K/sec %CPU K/sec %CPU K/sec %CPU
/sec %CPU
2000 33997 80.1 41932 22.9 11012 8.3 29599 83.4
95700 33.4 102.0 1.5

Its interesting to see that only in the random that the ata100 disk in
my box get wasted.

I think that in solaris 10 loging is on as default.
One question is what scsi control is being used in the sb1000 since it
is a fibre channel native, so scsi must be an add on card.