Re: parallel recovery slaves waiting on undo reads

From: Frits Hoogland <frits.hoogland_at_gmail.com>
Date: Sat, 29 Feb 2020 12:35:29 +0100
Message-Id: <B5F79869-1381-406F-A3A8-9C1FC6DB1B67_at_gmail.com>

Sadly, wait event timing can be influenced by database parameters, and recently I found that multi tenant changed the way the timing was done too.

Wait events typically (but not always!) time system call(s), for which the wait event time sometimes is an indication of performance in another layer in the application stack. I often use wait events to talk to for example storage admins about performance. Therefore, it’s critical that the wait event timings do correlate with timings of the admins of the other layer, so we can work together. This is one of the most important reasons I study wait events to the level that I do; so I understand what the timing incorporates, and therefore can explain that to for example the storage admin.

Last time I checked, the db file parallel read wait event timing for asynchronous IO looked like this:

1. io_submit (multiple IOs via an iocb struct, see: http://man7.org/linux/man-pages/man2/io_submit.2.html)
2. start wait event
3. io_getevents (blocking; wait for all IOs to finish)
4. end wait event

So not the total IO time is timed (although a very little part of it isn’t), and indeed if you disregard the small part that isn’t timed, it’s the timing of the slowest IO of all IOs that this wait event shows.

For synchronous IO, preadv() performs the same function, but with submission and waiting combined. All the IOs are submitted via an iovec (https://linux.die.net/man/2/preadv) using a single system call. The timing of this system call is obvious:

1. start wait event
2. preadv
3. end wait event

I can’t find a definitive source that tells me how preadv is implemented on linux. I would assume that linux is prepared for modern IO and does not assume it’s operating on a single disk and therefore performing the different IOs serially, but as I said, I would love to be pointed to the kernel source where the vector read is performed to validate it being serial or parallel. So for pread, I hope it’s the maximum time of the slowest IO, but it could be the sum of all individual IO times (serial).

I recently studied log file parallel write (again) for a conference is poland. Much to my surprise I found that the log file parallel write timing was done in the following way

1. io_submit
2. io_getevents (non-blocking; if all IOs are found goto 6)
3. start wait event
4. io_getevents (blocking)
5. end wait event
6. done

In other words: if the IO subsystem is fast enough, the wait event does not occur at all. This is consistent with what I found years ago with Oracle’s asynchronous direct path read implementation.

However, this was with multi-tenancy turned on. With it turned off, the timing became:

1. start wait event
2. io_submit
3. io_getevents
4. end wait event

I am surprised that multi-tenancy has this massive change in timing implementation. Of course the wait event timing (the latency) does not change that much, and both essentially give the IO time of the longest taking IO. In the light of new technologies like persistent writable memory (“pmem”) I can see this making sense: if the IO is nearly instantaneous, assume it is, and only start the time accounting (alias wait event) if it turns out it isn’t.

Frits Hoogland

http://fritshoogland.wordpress.com <http://fritshoogland.wordpress.com/> frits.hoogland_at_gmail.com <mailto:frits.hoogland_at_gmail.com> Mobile: +31 6 14180860

> On 28 Feb 2020, at 12:00, Noveljic Nenad <nenad.noveljic_at_vontobel.com> wrote:
>
> > From a performance perspective, the problem with this wait event is that the timing of the wait event has no absolute meaning: waiting for a single IO is something different that waiting for let’s say 70 IO requests submitted at the same time. p2 tells you the amount of oracle blocks, p3 the amount of requests.
>
> Assuming that ‘db file parallel read’ for multiple blocks starts measuring just before submitting IO operations and stops after the last wait completed, do you find the following interpretation of "db file parallel read" wait time correct when multiple blocks are involved:
> Async IO: “db file parallel read” wait time = max(IO time)
> Sync IO: “db file parallel read” wait time = sum(all IO times); (sum because the reads are executed sequentially, I think)
>
> Best regards,
>
> Nenad
>
> https://nenadnoveljic.com/blog <https://nenadnoveljic.com/blog>
>
>
> From: Frits Hoogland <frits.hoogland_at_gmail.com <mailto:frits.hoogland_at_gmail.com>>
> Sent: Freitag, 28. Februar 2020 09:53
> To: Noveljic Nenad <nenad.noveljic_at_vontobel.com <mailto:nenad.noveljic_at_vontobel.com>>
> Cc: Jonathan Lewis <jonathan_at_jlcomp.demon.co.uk <mailto:jonathan_at_jlcomp.demon.co.uk>>; oracle-l_at_freelists.org <mailto:oracle-l_at_freelists.org>
> Subject: Re: parallel recovery slaves waiting on undo reads
>
> Not sure how relevant this is, because you are looking for the reason your recovery worker processes do a lot of IO I assume.
>
> Below is a description of the what happens when you see db file parallel read:
>
> With newer oracle versions (12+), you’ll see plan lines indicating the word ‘BATCHED’. I believe it’s these points where oracle knows it has to read multiple non-adjacent blocks that it is getting these all at once. However, I read indications it might be happening outside of the ‘BATCHED’ lines, and is implemented at any time it knows multiple non-adjacent blocks are needed, which would have been read serially in the past.
>
> This can be implemented on the OS level as asynchronous IO requests via the regular asynchronous IO mechanism (io_submit-io_getevents), or uses a synchronous version to submit a a batch of IO requests: preadv. The mechanism of requesting multiple non-adjacent blocks has its own wait event: db file parallel read, which is a reasonable accurate description of what it actually does: it wants to read data from multiple places at the same time.
>
> From a performance perspective, the problem with this wait event is that the timing of the wait event has no absolute meaning: waiting for a single IO is something different that waiting for let’s say 70 IO requests submitted at the same time. p2 tells you the amount of oracle blocks, p3 the amount of requests.
>
> Frits Hoogland
>
> http://fritshoogland.wordpress.com <http://fritshoogland.wordpress.com/>
> frits.hoogland_at_gmail.com <mailto:frits.hoogland_at_gmail.com>
> Mobile: +31 6 14180860
>
>
>
> ____________________________________________________
> Please consider the environment before printing this e-mail.
> Bitte denken Sie an die Umwelt, bevor Sie dieses E-Mail drucken.
>
> Important Notice
>
> This message is intended only for the individual named. It may contain confidential or privileged information. If you are not the named addressee you should in particular not disseminate, distribute, modify or copy this e-mail. Please notify the sender immediately by e-mail, if you have received this message by mistake and delete it from your system.
> Without prejudice to any contractual agreements between you and us which shall prevail in any case, we take it as your authorization to correspond with you by e-mail if you send us messages by e-mail. However, we reserve the right not to execute orders and instructions transmitted by e-mail at any time and without further explanation.
> E-mail transmission may not be secure or error-free as information could be intercepted, corrupted, lost, destroyed, arrive late or incomplete. Also processing of incoming e-mails cannot be guaranteed. All liability of Vontobel Holding Ltd. and any of its affiliates (hereinafter collectively referred to as "Vontobel Group") for any damages resulting from e-mail use is excluded. You are advised that urgent and time sensitive messages should not be sent by e-mail and if verification is required please request a printed version.
> Please note that all e-mail communications to and from the Vontobel Group are subject to electronic storage and review by Vontobel Group. Unless stated to the contrary and without prejudice to any contractual agreements between you and Vontobel Group which shall prevail in any case, e-mail-communication is for informational purposes only and is not intended as an offer or solicitation for the purchase or sale of any financial instrument or as an official confirmation of any transaction.
> The legal basis for the processing of your personal data is the legitimate interest to develop a commercial relationship with you, as well as your consent to forward you commercial communications. You can exercise, at any time and under the terms established under current regulation, your rights. If you prefer not to receive any further communications, please contact your client relationship manager if you are a client of Vontobel Group or notify the sender. Please note for an exact reference to the affected group entity the corporate e-mail signature. For further information about data privacy at Vontobel Group please consult www.vontobel.com <https://www.vontobel.com/>.
>

--
http://www.freelists.org/webpage/oracle-l

Received on Sat Feb 29 2020 - 12:35:29 CET