Re: Exadata writeback Flash Cache usable space

From: Ls Cheng <exriscer_at_gmail.com>
Date: Mon, 19 Jul 2021 22:33:18 +0200
Message-ID: <CAJ2-Qb-sW11PV4VhePV4g94kqTsiJ-eVMDUS=qXTzL76gyQ1_w_at_mail.gmail.com>



Hi

I have to enable write back for flash cache in x9m because there is a write bottleneck in the current X6 Exadata where flash cache is configured with write though. So do we agree writeback reduces flash cache by the magnitude of ASM redundancy? By x2 (normal redundancy) or x3 (high redundancy). This was my main question.

Thanks

On Mon, Jul 19, 2021 at 1:28 PM Rajesh Aialavajjala < r.aialavajjala_at_gmail.com> wrote:

> I guess this is 'nuanced'...Flash cache mirroring vs ASM mirroring... Writeback
> usually reduces the effective size of the cache because multiple copies
> need to be retained to ensure that data integrity is maintained, even in
> the event of a failure.
>
> When write-back cache is enabled, the acknowledgement comes after the data
> is written to FlashCache. Unlike a DRAM cache, Flash is persistent
> storage, so it doesn't need to be written to disk. * To protect against
> failure of the Flash, all writes are multiplexed across either 2 or 3
> Exadata cells depending on the choice of NORMAL or HIGH redundancy
> respectively.*
>
> The data is NOT written to disk until there is space-pressure in the
> Flash.
>
> With Writeback flash cache, the data is redundant based on the defined
> redundancy for the diskgroup you are writing to. *ASM takes care of the
> redundancy independently of whether cellsrv writes the data to the flash or
> hard disk*.
>
> Given that this is an X9 - PMEM comes into play -- the default setting for
> PMEM Cache is Write-Through.
>
> Again -- additions, comments, corrections are welcome.
>
> Thanks,
>
> --Rajesh
>
> On Mon, Jul 19, 2021 at 3:48 AM Ls Cheng <exriscer_at_gmail.com> wrote:
>
>> Hi
>>
>> Since writeback can contain dirty data it needs mirroring, since the
>> dirty data in writeback mode contains database modified blocks I think it
>> should follow the same redundancy as the disk groups.
>>
>> I just found this yesterday, a bit dated but I think it applies
>>
>>
>> https://uhesse.com/2014/08/06/why-write-through-is-still-the-default-flash-cache-mode-on-exadata-x-4/
>>
>>
>> Thanks
>>
>>
>>
>> On Sun, Jul 18, 2021 at 3:07 PM Rajesh Aialavajjala <
>> r.aialavajjala_at_gmail.com> wrote:
>>
>>> I do not think that this is necessarily the case.
>>>
>>> The flashcache mode being set to "WriteBack" should not reduce the
>>> capacity of the flash cache available by a third. Also - I do not think
>>> that there's a dependency or link between the redundancy of the ASM disk
>>> groups and the flashcache mode.
>>>
>>> Each HC cell is provisioned with 25.6 TB of "raw" PCI flash...
>>>
>>> From an environment that has "WriteBack Flash Cache Enabled" - I see
>>> a reported cache size of 23 odd TB for each cell (this is an X8M-2)
>>>
>>> # cellcli -e list flashcache detail
>>> effectiveCacheSize: 23.28692626953125T
>>> id:
>>> size: 23.28692626953125T
>>> status: normal
>>>
>>> Additions, comments, corrections are welcome.
>>>
>>> Thanks,
>>>
>>> --Rajesh
>>>
>>> On Sat, Jul 17, 2021 at 4:22 AM Ls Cheng <exriscer_at_gmail.com> wrote:
>>>
>>>> Hi
>>>>
>>>> We are deploying a new Exadata x9m quarter rack with HC disks next week
>>>> and the flashcache will be configured using writeback instead of
>>>> writethrough.
>>>>
>>>> In writeback mode the flash is mirrored whereas in writethrough if I
>>>> understood correctly there is no mirroring. So I have a doubt, the quarter
>>>> rack has 76.8 TB raw flash space, what is the usable space with writeback
>>>> enabled? Since we are using High Redundancy I expect 25.6TB usable flash
>>>> cache space, correct?
>>>>
>>>> Thanks
>>>>
>>>>
>>>>

--
http://www.freelists.org/webpage/oracle-l
Received on Mon Jul 19 2021 - 22:33:18 CEST

Original text of this message