« MySQL Performance: Binlog Group Commit in 5.6 | Main | MySQL Performance: Table Open Cache in 5.6 »

Monday, 02 July, 2012

MySQL Performance: Read-Only Adventure in MySQL 5.6

It's from a log time now that I wanted to share the following stuff, but running time was always against me.. ;-)

I'd say that during last years there were plenty incredible events around MySQL Performance. And one of them is a story about Read-Only performance in MySQL 5.6 -- it's quite fun to tell it now, but, believe me, it's even more fun to live it yourself ;-)) Every important milestone around performance improvement in MySQL was similar to resolve a Detective Story ;-) Probably one day somebody will take a time to write a book about ;-)) But for my part, I'm just sharing what I've found interesting and important..

First of all: why even care about Read-Only performance in MySQL/InnoDB ?.. - Well, except if your database is "write-only", but usually data reads are representing a significant part in OLTP activity. And if your database engine is transactional, the reads from such a database are not "simple reads" -- they will also be a part of transactions, hitting different level constrains to match committed-only data or accept "dirty reads", etc. etc. etc. -- just to say that it's not as simple as it looks like ;-)) And on the same time read-only or read-dominated performance represents very significantly an overall database engine design, as it'll fire the most critical internal bottlenecks and give us an idea about its true scalability limits..

NOTE: I'm not covering here IO-bound workloads.. - it'll be already quite a long article, so I'm trying to make it shorter ;-)
(then, as you know, since you become IO-bound, your database performance is driven by your storage performance since then.. - and any DBA will simply advise you to buy a faster storage or more RAM ;-)) -- so, for the moment I'll worry about such a guy who has enough of RAM and still hitting performance/scalability issues ;-))

Split of the kernel_mutex

So far, it's more than a year ago there was shipped the very first working MySQL 5.6 version. It was in April 2011, MySQL 5.6.2 came with one of fundamental changes -- kernel_mutex contention was finally removed within InnoDB!.. This contention was seen as the main scalability stopper on many kind of workloads (including read-only). And Sunny made an incredible work to split kernel_mutex into several mutexes within InnoDB! The result of this work looked very promising. However, there is always a risk that by resolving one contention, you're also making another one more significant (which was simply hidden for the being time by the first one, which was more important until now).. - And it's exactly what happened in MySQL 5.6.2: instead of kernel_mutex contention we moved now to the TRX mutex contention (trx_sys), and in some particular cases performance became even worse than before.. You may find more details about in my MySQL 5.6 notes articles.

But well, the only reason for problems to exist is to be fixed, isn't it?.. ;-))

Introducing of READ-ONLY transactions

TRX mutex is mainly protecting data related to transactions within InnoDB code. While, when a part of your database activity is read-only, the question is coming in mind: is it possible to simplify little bit all this transactional overhead for a read-only operations?.. - And Sunny again came with a brilliant idea to introduce a new type of transactions within InnoDB - READ-ONLY TRANSACTIONS! - I think that his new feature was little bit missed within all other new improvements, while it was the first true and visible progress in MySQL 5.6 performance improvement! ;-))

How it works?..

  • if you're using auto-commit turned ON, there is even nothing to do, InnoDB will do all the stuff "automagically" for you! ;-)) -- if you're not using BEGIN/END/FOR UPDATE in your SELECT queries, they will be automatically turned into READ-ONLY transactions and most of transactional overhead will be avoided during your SELECT execution!
  • otherwise you may explicitly START TRANSACTION READ ONLY and get the same performance benefit for your SELECT queries!

For more details, please read the Sunny's article explaining all the related stuff and presenting impressive benchmark results on Sysbench!

By the end of 2011 year we've shipped MySQL 5.6.4 including all these new features, and MySQL 5.6 started to look better comparing to 5.5 even on my dbSTRESS tests ;-))

However, "better" is not enough.. What we're expecting from MySQL 5.6 is to be "way better" than any previous release ;-))

What's next?..

Then we entered into the most painful period, when you're asking yourself several times per day "do you really see what you see?.." and "do you even believe to what you see?.."

To get it by order:
  • since introduction of READ-ONLY transactions, near no other/new contention was reported by InnoDB during read-only workloads..
  • however, we were still unable to significantly out-pass 16 concurrent users sessions performance on servers with more than 16cores..
  • it was clear there is still a contention somewhere in the MySQL or InnoDB code..
  • and on this step we were really happy to use PERFORMANCE SCHEMA (PFS) as it's the only way to trace a time spent on every levels of MySQL code (particularly on mutexes and RW-locks in the given case)..
  • monitoring with PFS clearly indicated that the most of time is spent not on the LOCK_open mutex and MDL related locks!
  • by chance, Mikael already started to work with Dmitry Lenev to remove LOCK_open contention, but the final solution was not ready yet on that time..
  • to get an idea about potential gain in MySQL 5.6 performance once the LOCK_open mutex contention will be fixed, Dmitry made a "dirty patch" removing completely the use of LOCK_open and/or all MDL related stuff (just for testing propose only)..
  • And?.... ;-)

As you may imagine, I've expected to see performance way higher once both patches were applied... - but instead I've observed the following on a simple Sysbench OLTP_RO test:

Test case :
  • 32cores server
  • Sysbench OLTP_RO workload
  • each test series is executed with 16 and then with 32 concurrent users sessions (threads)
  • test series:
    • #1 -> MySQL 5.6 original
    • #2 -> #1 + patch removing any use of LOCK_open mutex
    • #3 -> #2 + patch removing any use of MDL related locks

Results :

Observations :
  • as you can see, within all test series QPS level is exactly the same!..
  • from the InnoDB mutex waits you may see that within all test series the main wait is on trx_sys mutex (but looking on the numbers you'll also see that it's a very low wait ;-)
  • however, according PFS, time spent on trx_sys mutex is quite important..
  • and then from PFS about test series:
    • #1 - both LOCK_open and MDL_lock times are present and they are top waited times
    • #2 - LOCK_open time is disappeared, while MDL_lock is still present (as expected)
    • #3 - both LOCK_open and MDL_lock times are gone.. - but performance still remained the same?.. - how it's possible?..

I'd admit that during this time I've even traced PFS code to see if what it's reporting is really matching the reality ;-))

All these and other problems we placed into agenda of our MySQL Performance meeting, which this year took place in Paris.. And this meeting was a full success. After one week of brainstorming and common efforts, we finally found the root problem of all observed bottlenecks.. ;-))

G5 patch

This is one of the cases when a database development is meeting HPC level problems ;-)) I've already dedicated a full article for the story of this remarkable improvement, will just note here that bottleneck was related to CPU cache line misses, and once fixed - performance was improved at least by 50% even on the less sensible HW, while on some workloads we even reached x6 times(!) better performance ;-))

The codename "G5" was assigned to this performance patch to reference the list of all improvement applied to MySQL 5.6 related to CPU caches. And since this critical point, our life finally moved to a normal way -- we moved back to true contentions as reported by InnoDB and Performance Schema ;-))

What's next?..

Of course, the most important improvement now was expected from the LOCK_open and MDL patch. However, once again, for our big surprises, removing any use of LOCK_open & MDL stuff did not bring any changes... Why?.. - a more detailed analyze with PFS showed that the most bigger amount of time was now again spent within TRX mutex contention! And it became quite clear, that until TRX mutex is not fixed, there will be no any benefit at all from LOCK_open and MDL fixes..

InnoDB spin wait delay setting is coming in the game

I'd say it was little bit frustrating to observe so high TRX contention even after so impressive speed-ups in already achieved performance levels.. But well, Sunny again, by analyzing code around mutex contention came with conclusion that increasing of wait delays during mutex spins will help to lower observed contentions! - Interesting that innodb_spin_wait_delay setting variable existed from a long date within InnoDB, but I never saw it helping to reduce any mutex contention before ;-) But now, since we came back to pure code contentions, this setting started to shine again!

I've spent an amount of time to analyze it, and to demonstrate the impact of this setting in MySQL 5.6 let me present you a short test result:

Test case:
  • Sysbench OLTP_RO workload
  • concurrent user sessions: 4, 8, 16, 32, 64, 128, 256
  • each load level is starting with innodb_spin_wait_delay=6
  • then, 3 minutes later, innodb_spin_wait_delay is set dynamically to 12
  • then to 24, 48 and 96

Observations :
  • first of all let's note a near linear scalability on the results up to 32 concurrent users! (I still remember the time when on 32cores performance was only worse, and on 32 users the result was always worse than on 16 ;-)) and the true changes came with MySQL 5.5, and then it's only continuing ;-))
  • then, since 64 users, the impact of innodb_spin_wait_delay (sd) become very visible:
    • with sd=6 we're blocked on 120K QPS
    • with sd=12 we're reaching 150-160K QPS
    • and finally with sd=96 just slightly more than 160K QPS
    • however, there is a huge gap between 120K and 160K QPS ;-))
    • and as you can see from the upper graph -- the regression is due increased contention on the TRX mutex with a growing workload..

NOTE: the default value of innodb_spin_wait_delay in MySQL 5.5 is 6, so it'll be changes since MySQL 5.6, and for sure it'll be bigger than 6 ;-))

But what exactly the meaning of this setting within InnoDB?..
  • when InnoDB is needing to acquire a mutex, the code will spin little bit if this mutex cannot be acquired at once (spin - means do several loops with short sleeps and try to acquire it again)
  • the "innodb_sync_spin_loops" setting is telling how many such loops should be involved (default: 30)
  • and the "innodb_spin_wait_delay" setting is simply defining how wide the range of the sleep interval should be used to choose (randomly) a value for sleeping (so, setting innodb_spin_wait_delay=96 doesn't mean that InnoDB will sleep 96us, but just a randomly chosen value between 0 and 96 microseconds)

This reminds me the discussion with Vadim last year about his observations on the innodb_sync_spin_loops impact while hitting kernel_mutex contention. And I also think that these spin related settings should be auto-adaptable and used per mutex rather globally. As well, according my observations, the spin delay setting is having not the same impact on mutexes as on RW-locks. But well, each thing on its time.. ;-))

LOCK_open contention removed

And only then, after all this improved stuff, we finally were able finally to see the benefit of the latest improvement in MySQL server code and made by Mikael and Dmitry Lenev -- split of LOCK_open mutex contention!.. This gave us yet another 10-15% performance improvement!

And then we're hitting the MDL locks contention.. well, as expected ;-))

So, as you see, MySQL 5.6 is coming with huge improvements in performance! However for the moment the "innodb_spin_wait_delay" is playing a significant role for InnoDB performance tuning.

Let's go little bit further now, and see what happens now to other stuff like AHI, and where we're now with an overall scalability limits..

Adaptive Hashing Index (AHI) and Spin Wait Delay impact

Keeping in mind we're hitting real contentions since now, the question which is coming back again is: use or not to use Adaptive Hash Index (AHI) ?.. - in many workloads its contention on the btr_search_latch is coming on the top position, however I'm not really sure that waits reported on it by InnoDB are the true waits.. Also, some are claiming to ss a better performance when AHI is turned off. Well, let's see the impact of AHI and its relation with spin wait delay (sd) on the following tests:

Test scenario:
  • workloads: dbSTRESS RO, Sysbench RO S-ranges, Sysbench OLTP_RO
  • server: 32cores Intel, 128GB RAM
  • concurrent users: 8, 16, 32, .. 512
  • configs:
    • ahi=on, sd=12 (spin delay)
    • ahi=on, sd=24
    • ahi=on, sd=96
    • ahi=off, sd=12
    • ahi=off, sd=24
    • ahi=off, sd=96

Observations :
  • with AHI=On the main contention is on the btr_search_latch RW-lock, and setting spin delay=12 or 24 is looking like the most optimal
  • with AHI=Off the main contention is moving to the Buffer Pool mutexes, and setting spin delay=96 is the most appropriate..
  • NOTE: interesting that with AHI disabled performance is still not better..

Sysbench RO S-ranges:

Observations :
  • setting spin delay=96 giving the best results for both AHI=On and AHI=Off
  • interesting that combination of AHI=On + spin delay=96 is giving the best result here
  • also, independently to AHI setting (On/Off), the main contention on this workload is remaining on the Buffer Pool mutex..

Sysbench OLTP_RO:

Observations :
  • setting spin delay to 24 or 96 is giving the best overall result with both AHI=On and AHI=Off
  • TRX sys mutex seems to be the main contention on this workload..

Analyzing Read-Only Scalability on 32cores Server

So, from the previous results, seems to me that for the scalability tests I have to use spin wait delay set to 24 when AHI=On, and 96 when AHI=Off. Let's get a look on results with AHI=On first.

Test scenario:
  • MySQL 5.6-m9 vs 5.6.4
  • setting: AHI=On, spin wait delay=24
  • workloads: dbSTRESS-RO, Sysbench RO S-ranges, Sysbench OLTP_RO
  • concurrent users: 8, 16, 32 .. 512
  • same 32cores server, but each test case is repeated with MySQL server bound via "taskset" on:
    • #1 -- all 32cores bi-thread (no binding, all 4 sockets are used)
    • #2 -- 32cores single-thread (binding on 4 sockets, one thread from each core)
    • #3 -- 16cores single-thread (binding on 2 sockets, one thread from each core)
    • #4 -- 8cores single-thread (binding on 1 socket, one thread from each core)

So, to keep it compact, on the following graphs you'll see 8 test series:
  • first 4 parts of graphs are corresponding to MySQL 5.6-m9 results within #1, #2, #3 and #4 config
  • then the same test series, but for MySQL 5.6.4..


Observations :
  • first surprise is to see the contention on "btr_search_latch" completely disappearing from InnoDB reporting once only one thread per core is used..
  • then, it's sure, there is still something going wrong, as on 16cores we have a better results vs 32cores.. - I'm surprised myself how often over a time I'm hitting various MySQL limits while testing dbSTRESS workload ;-))
  • at least the positive thing is that MySQL 5.6-m9 is out-passing 5.6.4 ;-)
  • well, dbSTRESS-RO should be yet more investigated..

Sysbench RO S-Ranges:
Observations :
  • these graphs are at least looking like I've expected ;-)
  • the best result on 5.6-m9 is when all 32cores with all core threads are used
  • having a second thread on CPU core is helping in 5.6-m9
  • scalability is nor perfect on 5.6-m9, but performance is constantly growing with more cores available ;-)
  • while on 5.6.4 the best result is reached while only one CPU socket is used (hitting the cache line issue fixed within April's 5.6-labs)..
  • up to x6 times performance improvement in 5.6-m9, which is really good ;-)

Sysbench OLTP_RO:

Observations :
  • 5.6-m9 performance on 32cores is better than on 32cores bi-thread, seems that TRX sys mutex contention is increasing in this case..
  • otherwise, pretty good scalability on 5.6-m9 while moving from 8 to 16 and 32cores
  • while 5.6.4 is getting a better performance on 16 vs 8cores, but loosing it on 32cores again (hitting here CPU cache line issue + other internal contentions)..
  • up to x3 times better performance on 5.6-m9 vs 5.6.4 ;-))

The same tests now, but with AHI=Off and spin delay=96

NOTE: Interesting that results with disabled AHI are quite similar to those with AHI=On, but performance is still better when AHI is turned ON. And, yes, it's clear I'm hitting another "hidden bottleneck" within MySQL 5.6 on dbSTRESS workload..

dbSTRESS-RO, AHI=Off, sd=96:

Sysbench RO S-ranges, AHI=Off, sd=96:

Sysbench OLTP_RO, AHI=Off, sd=96:

What else to add?..
Just express my personal kudos to Sunny, Mikael and Dmitry! and the whole MySQL Team! ;-))

Then, work is continuing, and new challenges are already waiting for us.. ;-))
As usually, comments are welcome!

Posted by Dimitri at 16:23
Categories: MySQL
blog comments powered by Disqus
Note: if you don't see any "comment" dialog above, try to access this page with another web browser, or google for known issues on your browser and DISQUS..