Re: Weak disk I/O performance on daX compared to adaX, was: Re: dd performance [was Re: Call for Foundation-supported Project Ideas]

On Mon, Dec 13, 2021 at 8:08 PM Stefan Blachmann <sblachmann@gmail.com>
wrote:

> I am wondering what could be the cause for the weak disk I/O
> performance on FreeBSD when using daX drivers instead of adaX drivers.
>
> Explanation:
> The HP Z420 has 6 SATA ports.
> SATA drives that get connected to port #1 to #4 are being shown as daX
> drives on FreeBSD.
> Drives connected to ports 5 and 6 appear as adaX drives.
>
> > On 12/2/21, Alan Somers <asomers@freebsd.org> wrote:
> >> That is your problem then.  The default value for dd if 512B.  If it
> >> took 3 days to erase a 2 TB HDD, that means you were writing 15,000
> >> IOPs.  Frankly, I'm impressed that the SATA bus could handle that
>
> This shows that on the ada driver, the disk I/O performance is acceptable.
> However, after 14 days dd is still working on the same type drive on
> connector 4 (da3).
>
> So my questions:
> - Why does FreeBSD use the da driver instead of the ada driver for
> drives on SATA ports 1-4?
>

Because it isn't an ahci controller, but something else. dmesg will tell
you, or camcontrol devlist. It might be crappy hardware, poorly configured
or maybe you've discovered a driver bug that's easy to fix.


> - And, why is the da driver so slow? (For example, on HP Z800 when
> used with FreeBSD, 15k SAS drives seem as slow as normal consumer
> drives, while on Linux disk I/O is just snappy.)
>

It isn't. It's more likely the controller they are attached to that's slow.
At work
we get line rate out of daX and adaX all the time. They are just protocol
translators and hands it off to the host adapter (what's called the SIM).


> - Is there a way to configure FreeBSD to use the ada driver instead of
> the da driver, so using FreeBSD is still an alternative to Linux if
> disk speed matters?
>

Unlikely.


> - Or is it impossible to use the ada drivers on SATA connectors 1-4
> for maybe some HP Z420 hardware related reasons?
>

What does camcontrol devlist tell you? Chances are it's the SIM that's to
blame for the poor performance (we run all kinds of crazy I/O through ada
and da and if anything da is a smidge faster). The key question is why
things are so seemingly slow.

Warner


> Cheers,
> Stefan
>
>
> On 12/2/21, Stefan Blachmann <sblachmann@gmail.com> wrote:
> > Ah, the buffer cache! Didn't think of that.
> > Top shows the weighted cpu load is about 4%, so your guess that it was
> > the SATA scheduler might be correct.
> > Will try this on Linux the next days using conv=direct with a pair of
> > identical HDDs.
> > Already curious for the results.
> >
> >
> >
> > On 12/2/21, Alan Somers <asomers@freebsd.org> wrote:
> >> That is your problem then.  The default value for dd if 512B.  If it
> >> took 3 days to erase a 2 TB HDD, that means you were writing 15,000
> >> IOPs.  Frankly, I'm impressed that the SATA bus could handle that
> >> many.  By using such a small block size, you were doing an excellent
> >> job of exercising the SATA bus and the HDD's host interface, but its
> >> servo and write head were mostly just idle.
> >>
> >> The reason why Linux is different is because unlike FreeBSD it has a
> >> buffer cache.  Even though dd was writing with 512B blocks, those
> >> writes probably got combined by the buffer cache before going to SATA.
> >> However, if you use the conv=direct option with dd, then they probably
> >> won't be combined.  I haven't tested this; it's just a guess.  You can
> >> probably verify using iostat.
> >>
> >> When you were trying to erase two HDDs concurrently but only one was
> >> getting all of the IOPs and CPU time, was your CPU saturated?  I'm
> >> guessing not.  On my machine, with a similar HDD, dd only consumes 10%
> >> of the CPU when I write zeros with a 512B block size.  I need to use a
> >> 16k block size or larger to get the IOPs under 10,000.  So I'm
> >> guessing that in your case the CPU scheduler was working just fine,
> >> but the SATA bus was saturated, and the SATA scheduler was the source
> >> of the unfairness.
> >> -Alan
> >>
> >> On Thu, Dec 2, 2021 at 10:37 AM Stefan Blachmann <sblachmann@gmail.com>
> >> wrote:
> >>>
> >>> I intentionally used dd without the bs parameter, as I do care less
> >>> about "maximum speed" than clearing the drives completely and also do
> >>> a lot of I/O transactions.
> >>> The latter because drives that are becoming unreliable tend to
> >>> occasionally throw errors, and the more I/O transactions one does the
> >>> better the chance is to spot this kind of drives.
> >>>
> >>> The system is a HP Z420, the mainboard/chipset/controller specs can be
> >>> found in the web.
> >>> The drives in question here (quite old) 2TB WD Black enterprise grade
> >>> 3.5" SATA drives. Their SMART data is good, not hinting at any
> >>> problems.
> >>>
> >>> On Linux, erasing them both concurrently finished at almost the same
> >>> time.
> >>> Thus I do not really understand why on FreeBSD this is so much
> >>> different.
> >>>
> >>> On 12/2/21, Alan Somers <asomers@freebsd.org> wrote:
> >>> > This is very surprising to me.  I never see dd take significant CPU
> >>> > consumption until the speed gets up into the GB/s range.  What are
> you
> >>> > using for the bs= option?  If you set that too low, or use the
> >>> > default, it will needlessly consume extra CPU and IOPs.  I usually
> set
> >>> > it to 1m for this kind of usage.  And what kind of HDDs are these,
> >>> > connected to what kind of controller?
> >>> >
> >>> > On Thu, Dec 2, 2021 at 9:54 AM Stefan Blachmann <
> sblachmann@gmail.com>
> >>> > wrote:
> >>> >>
> >>> >> Regarding the suggestions to either improve or replace the ULE
> >>> >> scheduler, I would like to share another observation.
> >>> >>
> >>> >> Usually when I need to zero out HDDs using dd, I use a live Linux.
> >>> >> This time I did that on FreeBSD (13).
> >>> >> My observations:
> >>> >> - On the same hardware, the data transfer rate is a small fraction
> >>> >> (about 1/4th) of which is achieved by Linux.
> >>> >> - The first dd process, which erases the first HDD, gets almost all
> >>> >> CPU and I/O time. The second process which does the second HDD is
> >>> >> getting starved. It actually really starts only after the first one
> >>> >> finished.
> >>> >>
> >>> >> To me it was *very* surprising to find out that, while erasing two
> >>> >> similar HDDs concurrently takes about one day on Linux, on FreeBSD,
> >>> >> the first HDD was finished after three days, and only after that the
> >>> >> remaining second dd process got the same CPU time, making it proceed
> >>> >> fast instead of creepingly slowly.
> >>> >>
> >>> >> So I guess this might be a scheduler issue.
> >>> >> I certainly will do some tests using the old scheduler when I got
> >>> >> time.
> >>> >> And, I ask myself:
> >>> >> Could it be a good idea to sponsor porting the Dragonfly scheduler
> to
> >>> >> FreeBSD?
> >>> >>
> >>> >> On 12/2/21, Johannes Totz <jo@bruelltuete.com> wrote:
> >>> >> > On 29/11/2021 03:17, Ed Maste wrote:
> >>> >> >> On Sun, 28 Nov 2021 at 19:37, Steve Kargl
> >>> >> >> <sgk@troutmask.apl.washington.edu> wrote:
> >>> >> >>>
> >>> >> >>> It's certainly not the latest and greatest,
> >>> >> >>> CPU: Intel(R) Core(TM)2 Duo CPU     T7250  @ 2.00GHz
> (1995.04-MHz
> >>> >> >>> K8-class CPU)
> >>> >> >>
> >>> >> >> If you're content to use a compiler from a package you can save a
> >>> >> >> lot
> >>> >> >> of time by building with `CROSS_TOOLCHAIN=llvm13` and
> >>> >> >> `WITHOUT_TOOLCHAIN=yes`. Or, instead of WITHOUT_TOOLCHAIN perhaps
> >>> >> >> `WITHOUT_CLANG=yes`, `WITHOUT_LLD=yes` and `WITHOUT_LLDB=yes`.
> >>> >> >
> >>> >> > (re-send to list, sorry)
> >>> >> > Can we disconnect the compiler optimisation flag for base and
> >>> >> > clang?
> >>> >> > I
> >>> >> > don't need the compiler to be build with -O2 but I want the
> >>> >> > resulting
> >>> >> > base system to have optimisations enabled.
> >>> >> > Right now, looks like both get -O2 and a lot of time is spent on
> >>> >> > optimising the compiler (for no good reason).
> >>> >> >
> >>> >> >
> >>> >>
> >>> >
> >>
> >
>
>