Re: File system performance, hardware performance, ext3, 3ware RAID1, etc.

[Timothy Miller]

Daniel Blueman wrote:
> Tim,
>
> Do you get the same numbers (but slightly higher, as this is will measure
> from a smaller portion of outer zones) with:
>
> # hdparm -t /dev/sda
>
> ?

I ran this test.  This is a read test.  What I did below was a write test.

Additionally, I ran this test:
    time dd if=/dev/sda of=/dev/null bs=1024k count=1024

From that, I got 47 megs/sec.  From 'hdparm -t /dev/sda', I got a 
slightly lower number.

So, for reads, I'm getting good performance.  47 megs/sec at the 
outer-most tracks is a bit lower than the 50+ that reviewers report, but 
it's not bad.

However, I don't get anywhere near the 40+ megs/sec the reviewers say 
the drive gets for writes.  That model as a single drive in my wife's 
computer gets about 39 megs/sec, which is great.  But behind the 3ware, 
the drive gets only 13 megs/sec.  (iozone reports about 15 megs/sec, but 
that's influenced by caching in RAM, and iozone is writing to a file on 
tracks further out, I think.)

> > Adam Radford wrote:
> >
> > > Perhaps you are issuing non purely sequential IO.  The card firmware
> >
> > does
> >
> > > some 
> > > reodering, but at some point it will cause performance degradation.  Can
> >
> > you
> >
> > > try 
> > > kernel 2.6 w/xfs? 
> >
> > Not any time soon, but as I mentioned earlier, I measured 13.9 megs/sec 
> > when I ran this command:
> >
> >     time dd if=/dev/zero of=/dev/sda2 bs=1024k count=1024
> >
> > No file system was involved; I was simply writing zeros to the block 
> > device (swap partition with swap off).  It took 73.522 seconds to do the 
> > above operation.  Also, I was running in single-user mode while doing 
> > the test.
> >
> >
> > > Also, in my experience, the 'raw io' interface doesn't issue any
> > > asynchronous IO.  The
> > > card _definately_ needs asynchronous IO posted to it or you will not get
> > > good results
> > > because you won't get all the drives busy.
> >
> > With RAID1, both drives will be written with the same data.  There is no 
> > need to be asynchronous, since it's all completely linear and sequential 
> > with large data blocks.

-
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/