Skip site navigation (1)Skip section navigation (2)
Date:      Wed, 26 Oct 2005 10:00:45 +0100
From:      Joao Barros <joao.barros@gmail.com>
To:        Robert Atkinson <phreaki@gmail.com>
Cc:        freebsd-current@freebsd.org
Subject:   Re: Poor Samba throughput on 6.0 RC1
Message-ID:  <70e8236f0510260200r60ba38abha0a3afefcfa976d4@mail.gmail.com>
In-Reply-To: <6fb2b4650510252107n66318924m874d530fb01410ad@mail.gmail.com>
References:  <70e8236f0510241518x7b280938jd15f7e8c3224cbd@mail.gmail.com> <435D64B2.2020703@linkline.com> <6fb2b4650510242021l72a1ceb9m91e72a0420458982@mail.gmail.com> <70e8236f0510251636k1002cc96yd357492a138e1933@mail.gmail.com> <6fb2b4650510252107n66318924m874d530fb01410ad@mail.gmail.com>

next in thread | previous in thread | raw e-mail | index | archive | help
On 10/26/05, Robert Atkinson <phreaki@gmail.com> wrote:
> On 10/25/05, Joao Barros <joao.barros@gmail.com> wrote:
> > On 10/25/05, Robert Atkinson <phreaki@gmail.com> wrote:
> > > Maybe a longshot, but what is your cluster size?
> > The default block size of 16384 bytes, a fragment size of 2048 bytes
> > The ide drive has a NTFS partition which I mounted readonly and
> > copying files from there resulted with the same below expected
> > performance :(
> >
> > >
> > >  I'd be interested to hear if same irq is being shared if that is eno=
ugh
> bw?
> >
> >
> > > I disabled the USB controller and the extra NIC to get more available
> > > IRQs and tried using the amr on diferent pci slots with no change in
> > > performance.
> >
>
>
> I could swear I remember seeing that actually usbd_enable=3D"NO" is requi=
red
> even
> if you disable the usb controller for the problem with drive speed and th=
e
> usbd.

I don't have usbd enabled.

>
> Above that, does dmesg still show the nic, agp, and scsi controller sitti=
ng
> on irq 11?
> Might take a reset config data in bios, or a staggering of pci slots to g=
et
> it to stop
> doing that.

No, they don't share interrupts. And I did the slot juggling ;)

>
>
> I'm not great at vm usage, but I think it could just be chipset issue.
> I've seen some really poor results with disk performance from some people=
,
> but was mostly related to vnconfig and mdconfig (4.11 vs 5.X).

For what it's worth, it's a server motherboard: Supermicro 370SSR

>
> This may not help your system out that greatly, as it doesn't appear to b=
e
> an issue,
> but your fxp card should be supported for device_polling. AFAIK you shoul=
d
> see better
> performance, since the fxp card won't be waking up the kernel for every
> little packet.
>

Polling should help when you have lots of interrupts which I believe I
do. My objective now is discovering why I do ;)

> I'd hope that with screen updates, nic data and the load of the irq from =
the
> disk would
> hurt it all. Surely the agp card is using up some of that, but I'm not su=
re
> of what that chipset would do under those types of conditions.

I mostly use ssh on that machine.

>
> Is dma being enabled? Is it really important to have access times marked =
on
> disk?
> You could/should turn that off with tunefs I think if you don't need it,
> will waste some
> access time and if you don't want that stamp, things should improve for a=
t
> least smaller files.

My test is with big files. More easy to benchmark.

>
> Above all, when a drive starts to be unusually slow, I'd pop a 4.11 disc =
in
> and test, throw in hard drive
> tester if 4.11 shows same perf, which is not easy for scsi drives (don't =
see
> smart data on the ones I have), ide drives can at least show me what erro=
rs
> are coming from the disk reads and I know the drive is going to be slower=
.
> (It takes 5 attempts to get a seek to go through)
>

I'm getting the same results from an IDE drive, a single SCSI-2 and a
RAID5 with 4 drives.
I take it it's not a drive issue.

Thanks for your input :)


--
Joao Barros



Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?70e8236f0510260200r60ba38abha0a3afefcfa976d4>