Re: A beginners guide to implementing ASM

From: Andre Maasikas <amaasikas_at_gmail.com>
Date: Thu, 4 Mar 2021 09:31:16 +0200
Message-ID: <CAL5UiscGgcdtOdmyfjc=RyR_RCvSOo3vk08cqz_y14qJbRj6gg_at_mail.gmail.com>


You probably are referring to your blog post and tests a while ago where you were using virtualbox with disks presented to guest via ATA controller which present them
with very unusual combination of write cache and no FUA support. This was my reply to you directly back then:

As getting posting privs takes time..
My current assumption is that the difference is because of the interaction of Oracle device access (O_DIRECT) and specific ATA drive/driver interaction, Because of O_DIRECT essentially all write requests have to be flushed to disk before
indication completion to Oracle.

As the ATA drives report write cache (at least in case of vbox and it doesn't allow to turn it off)
and no FUA support, linux kernel has to issue a CACHE FLUSH after every write request (or maybe even more) - see https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/Documentation/block/barrier.txt and seems it cannot queue more commands inbetween.

So I'v also compiled a custom kernel where I forcibly indicated 'no write cache' for ATA drives and
it brought the performance to the SCSI driver level.

Given that i'm not even sure that the asmfd does 'the right thing' given write cache enabled ata drives.
It might cheat/forget that one has to issue the flushes. I dont plan to investigate this further as none of my systems have write heavy load or SATA drives :)
but it was on interesting issue to analyze anyway :)

Andre

On Thu, Mar 4, 2021 at 3:41 AM Jared Still <jkstill_at_gmail.com> wrote:
>
>
> On Fri, Feb 26, 2021 at 7:07 PM Mladen Gogala <gogala.mladen_at_gmail.com> wrote:
>>
>> I have to say that I prefer the Linux native methods. They are a pain in
>> the neck (or lower) to set up but I haven't noticed any performance
>> advantages of the filter driver over the native method, set up by udev
>> rules.
>>
>
> We will have to agree to disagree on that point. I have measured it, and FD is faster than udev rules or ASMLib
>
> Jared
>

--
http://www.freelists.org/webpage/oracle-l
Received on Thu Mar 04 2021 - 08:31:16 CET

Original text of this message