Linux and H/W optimizations for MySQL

Linux and H/W optimizations for MySQL

View more presentations from Yoshinori Matsunobu

Linux and H/W optimizations for MySQL Yoshinori Matsunobu Principal Infrastructure Architect, DeNA Former APAC Lead MySQL Consultant at MySQL/Sun/Oracle Yoshinori.Matsunobu@dena.jp
Table of contents MySQL performance overview and improvement history SSD and MySQL performance Memory and swap space management File I/O Network Useful command line tools and tips
History of MySQL performance improvements H/W improvements HDD RAID, Write Cache Large RAM SATA SSD、PCI-Express SSD More number of CPU cores Faster Network S/W improvements Improved algorithm (i/o scheduling, swap control, etc) Much better concurrency Avoiding stalls Improved space efficiency (compression, etc)
Per-server performance is important To handle 1 million queries per second.. 1000 queries/sec per server : 1000 servers in total 10000 queries/sec per server : 100 servers in total Additional 900 servers will cost 10M$ initially, 1M$ every year If you can increase per server throughput, you can reduce the total number of servers, which will decrease TCO Sharding is not everything
32bit LinuxUpdates 2GB RAM 2GB RAM 2GB RAM HDD RAID HDD RAID HDD RAID (20GB) (20GB) (20GB) + Many slaves + Many slaves + Many slaves Random disk i/o speed (IOPS) on HDD is very slow 100-200/sec per drive Database easily became disk i/o bound, regardless of disk size Applications could not handle large data (i.e. 30GB+ per server) Lots of database servers were needed Per server traffic was not so high because both the number of users and data volume per server were not so high Backup and restore completed in short time MyISAM was widely used because it’s very space efficient and fast
64bit Linux + large RAM + BBWC 16GB RAM + Many slaves HDD RAID (120GB) Memory pricing went down, and 64bit Linux went mature It became common to deploy 16GB or more RAM on a single linux machine Memory hit ratio increased, much larger data could be stored The number of database servers decreased (consolidated) Per server traffic increased (the number of users per server increased) “Transaction commit” overheads were extremely reduced thanks to battery backed up write cache From database point of view, InnoDB became faster than MyISAM (row level locks, etc) Direct I/O became common
Side effect caused by fast server After 16-32GB RAM became common, we could run many more users and data per server Write traffic per server also increasedMaster 4-8 RAID 5/10 also became common, which improved concurrency a lot On 6 HDD RAID 10, single thread IOPS is around HDD RAID 200, 100 threads IOPS is around 1000-2000 Good parallelism on both reads and writes on master On slaves, there is only one writer thread (SQL thread). No parallelism on writes 6 HDD RAID10 is as slow as single HDD for writesSlave Slaves became performance bottleneck earlier than HDD RAID master Serious replication delay happened (10+ minutes at peak time)
Using SATA SSD on slaves IOPS differences between master (1000+) and slave (100+) have caused serious replication delay Is there any way to gain high enough IOPS from single thread?Master Read IOPS on SATA SSD is 3000+, which should be enough (15 times better than HDD) HDD RAID Just replacing HDD with SSD solved replication delay Overall read throughput became much better Using SSD on master was still risky Using SSD on slaves (IOPS: 100+ -> 3000+) was more effective than using on master (IOPS: 1000+ -> 3000+)Slave We mainly deployed SSD on slaves SATA SSD The number of slaves could be reduced From MySQL point of view: Good concurrency on HDD RAID has been required : InnoDB Pluguin
Concurrency improvements around MySQL Row level locks, not table/page level locks InnoDB: Row level locks, MyISAM: Table level locks No Hot Giant Lock(Global Mutex) Critical sections (spots that only single thread can be accessible) reduce concurrency a lot InnoDB Buffer Pool, Data Dictionary, etc InnoDB now scales pretty well with many CPU cores (5.1 Plugin, 5.5) Parallel I/O threads Increasing the number of disk i/o threads – Front-end operations (i.e SELECT) throughput improved Introducing asynchronous I/O Parallel background tasks Dividing checkpoint thread (main thread) and purge thread (physically removing delete-marked records) Single-threaded replication channel (SQL Thread) is currently hot spot
Avoiding sudden performance dropsPerf Product A Product B Time Some unstable database servers suddenly drop performance in some situations Low performance is a problem because we can’t meet customers’ demands Through product A is better on average, product B is much more stable Don’t trust benchmarks. Vendors’ benchmarks show the best score but don’t show worse numbers
Avoiding stalls All clients are blocked for a short period of time (less than one second – a few seconds) The number of connections grow significantly (10-100 on average, but suddenly grows to 1000+ and TOO MANY CONNECTIONS errors are thrown) Increased response time
Avoiding stalls(2) Mainly caused by database internal problems Some of them are known issues. Follow best practices Improved every day Don’t use too old MySQL versions (4.0/4.1, 5.0.45, etc..) Typical stalls Dropping a huge table (LOCK_open mutex) Burst write (at checkpoint, at the time when redo log file free space becomes not enough) pthread_create()/clone() (called at connection establishment) etc
Handling Real-World workloads.. Company Introduction: DeNA and Mobage Platform One of the largest social game providers in Japan – Both social game platform and social games themselves – Subsidiary ngmoco:) in SF Japan localized phone, Smart Phone, and PC games 2-3 billion page views per day 25+ million users 1.3B$ revenue in 2010
RDBMS or NoSQL ? Is MySQL good for social games ? It’s good! DeNA uses MySQL for data stores, memcached for caching summary/objects H/W is getting fast enough for RDBMS to handle lots of workloads (Nehalem CPU, Large RAM, PCI-Express SSD) Complex query, multiple columns, transaction, secondary index, online backup, monitoring, utility tools are still very helpful for social games Is RDBMS slow (Can’t MySQL handle 1000 updates/second) ? Some of our MySQL servers handle 10,000+ UPDATE statements per second Highly depending on random disk i/o speed and indexing Any NoSQL database can’t be fast if it becomes disk i/o bound Schema, transaction and indexes should be more taken care
Social Game workloads Easily increasing millions of users in a few days Database size grows rapidly – Especially if PK is “user_id + xxx_id” – Increasing GB/day is typical Load balancing reads is not difficult Adding slaves or caching servers Load balancing writes is not trivial Sharding Solutions depend on what kinds of tables we’re using, INSERT/UPDATE/DELETE workloads, etc
INSERT-mostly tables History tables such as access logs, diary, battle history INSERT and SELECT mostly Table size becomes huge (easily exceeding 1TB) Locality (Most of SELECT go to recent data) INSERT performance in general Fast in InnoDB (Thanks to Insert Buffering. Much faster than MyISAM) To modify index leaf blocks, they have to be in buffer pool When index size becomes too large to fit in the buffer pool, disk reads happen In-memory workloads -> disk-bound workloads – Suddenly suffering from serious performance slowdown – UPDATE/DELETE/SELECT also getting much slower Any faster storage devices can not compete with in-memory workloads
InnoDB Feature: Insert Buffering If non-unique, secondary index blocks are not in memory, InnoDB inserts entries to a special buffer(“insert buffer”) to avoid random disk i/o operations Insert buffer is allocated on both memory and innodb SYSTEM tablespace Periodically, the insert buffer is merged into the secondary index trees in the database (“merge”) Insert buffer Pros: Reducing I/O overhead Optimized i/o Reducing the number of disk i/o operations by merging i/o requests to the same block Some random i/o operations can be sequential Cons: Additional operations are added Merging might take a very long time when many secondary indexes must be updated and many rows have been inserted. it may continue to happen after a server shutdown and restart
INSERT gets slower Time to insert 1 million records (InnoDB, HDD) 600 500 2,000 rows/s Seconds 400 Sequential order 300 Random order 200 100 10,000 rows/s 0 1 13 25 37 49 61 73 85 97 109 121 133 145 Existing records (millions) Index size exceeded buffer pool size Index size exceeded innodb buffer pool size at 73 million records for random order test Gradually taking more time because buffer pool hit ratio is getting worse (more random disk reads are needed) For sequential order inserts, insertion time did not change. No random reads/writes
INSERT performance difference In-memory INSERT throughput 15000+ insert/s from single thread on recent H/W Exceeding buffer pool, starting disk reads Degrading to 2000-4000 insert/s on HDD, single thread 6000-8000 insert/s on multi-threaded workloads Serious replication delay often happens Faster storage does not solve everything At most 5000 insert/s on fastest SSDs such as tachIOn/FusionIO – InnoDB actually uses CPU resources quite a lot for disk i/o bound inserts (i.e. calculating checksum, malloc/free) It is important to minimize index size so that INSERT can complete in memory
Approach to complete INSERT in memory Partition 1 Partition 2 Single big physical table(index) Partition 3 Partition 4 Range partition by datetime Started from MySQL 5.1 Index size per partition becomes total_index_size / number_of_partitions INT or TIMESTAMP enables hourly based partitions – TIMESTAMP does not support partition pruning Old partitions can be dropped by ALTER TABLE .. DROP PARTITION
Approaches to complete INSERT in memory Dividing tables by datetime (up to 5.0, or if your tables are too big to store on single server’s disks) Purging or archiving older data Dropping indexes Dropping columns Using space efficient data types, compression Using large RAM Sharding
UPDATE-mostly tables Typical usage Updating users’ status – Current HP, experiences, money, number of items, friends’ status, battle status, etc – UPDATE and SELECT are most frequently executed for handling user status UPDATE performance Need to read target record blocks and index blocks – Fast if completed in buffer pool, otherwise massive foreground disk reads happen – Data size does not grow significantly, depending on INSERT/DELETE Huge performance difference between storage devices – In-memory UPDATE: 12,000/s – HDD UPDATE: 300/s – SATA SSD UPDATE: 1,800/s – PCI-E SSD UPDATE: 4,000/s – * Single Thread – Random reads happen in foreground, so random read speed matters a lot
SSD performance and deployment strategies for MySQL
What do you need to consider? (H/W layer) SSD or HDD? Interface SATA/SAS or PCI-Express?, How many drives? RAID H/W RAID, S/W RAID or JBOD? Network Is 100Mbps or 1Gbps enough? Memory Is 2GB RAM + PCI-E SSD faster than 64GB RAM + 8HDDs? CPU Nehalem, Opteron or older Xeon?
What do you need to consider? Redundancy RAID DRBD (network mirroring) Semi-Sync MySQL Replication Async MySQL Replication Filesystem ext3, xfs, raw device ? File location Data file, Redo log file, etc SSD specific issues Write performance deterioration Write endurance
Why SSD? IOPS! IOPS: Number of (random) disk i/o operations per second Almost all database operations require random access Selecting records by index scan Updating records Deleting records Modifying indexes Regular SAS HDD : 200 iops per drive (disk seek & rotation is slow) SSD : 2,000+ (writes) / 5,000+ (reads) per drive highly depending on SSDs and device drivers
Table of contents Basic Performance on SSD/HDD Random Reads Random Writes Sequential Reads Sequential Writes fsync() speed Filesystem difference IOPS and I/O unit size SLC vs MLC # of interfaces/drives Opteron vs Nehalem MySQL Deployments for benchmarks MySQL Deployments for real workloads
Random Read benchmark Direct Random Read IOPS (Single Drive, 16KB, xfs) 45000 40000 35000 30000 25000 HDD IOPS 20000 Intel SSD 15000 Fusion I/O 10000 5000 0 1 2 3 4 5 6 8 10 15 20 30 40 50 100 200 # of I/O threads HDD: 196 reads/s at 1 i/o thread, 443 reads/s at 100 i/o threads Intel : 3508 reads/s at 1 i/o thread, 14538 reads/s at 100 i/o threads Fusion I/O : 10526 reads/s at 1 i/o thread, 41379 reads/s at 100 i/o threads Single thread throughput on Intel is 16x better than on HDD, Fusion is 25x better SSD’s concurrency (4x) is much better than HDD’s (2.2x) Very strong reason to use SSD
High Concurrency Single SSD drive has multiple NAND Flash Memory chips (i.e. 40 x 4GB Flash Memory = 160GB) Highly depending on I/O controller and Applications Single threaded application can not gain concurrency advantage
PCI-Express SSD CPU North Bridge South Bridge PCI-Express Controller SAS/SATA Controller 2GB/s (PCI-Express x 8) 300MB/s SSD I/O Controller SSD I/O Controller Flash Flash Advantage PCI-Express is much faster interface than SAS/SATA (current) Disadvantages Most motherboards have limited # of PCI-E slots No hot swap mechanism
Write performance on SSD Random Write IOPS (16KB Blocks) 20000 18000 16000 14000 12000 1 i/o thread 10000 100 i/o threads 8000 6000 4000 2000 0 HDD(4 RAID10 xfs) Intel(xfs) Fusion (xfs) Very strong reason to use SSD But wait.. Can we get a high write throughput *everytime*? Not always.. Let’s check how data is written to Flash Memory
Understanding how data is written to SSD (1) Block (empty) Block (empty) Block (empty) Block Page Page . Flash memory chips Single SSD drive consists of many flash memory chips (i.e. 2GB) A flash memory chip internally consists of many blocks (i.e. 512KB) A block internally consists of many pages (i.e. 4KB) It is *not* possible to overwrite to a non-empty block Reading from pages is possible Writing to pages in an empty block is possible Appending is possible Overwriting to pages in a non-empty block is *not* possible
Understanding how data is written to SSD (2) Block (empty) Block (empty) New data Block (empty) Block × Page Page . Overwriting to a non-empty block is not possible Writing new data to an empty block instead Writing to a non-empty block is fast (-200 microseconds) Even though applications write to same positions in same files (i.e. InnoDB Log File), written pages/blocks are distributed (Wear-Leveling)
Understanding how data is written to SSD (3) Block P Block P Block P P P P 1. Reading all pages Block P Block P Block P P P New P 2. Erasing the block Block Block P Block P P P P 3. Writing all data P P In the long run, almost all blocks will be fully used New P i.e. Allocating 158GB files on 160GB SSD New empty block must be allocated on writes Basic steps to write new data: 1. Reading all pages from a block 2. ERASE the block 3. Writing all data w/ new data into the block ERASE is very expensive operation (takes a few milliseconds) At this stage, write performance becomes very slow because of massive ERASE operations
Reserved SpaceData Space Reserved Space Block P Block P Block P Block (empty) P P P Block P Block P Block P Block (empty) P P P Block Block P Block P 2. Writing data P P P 1. Reading pages P New dataBackground jobs ERASE unused blocks P To keep high enough write performance, SSDs have a feature of “reserved space” Data size visible to applications is limited to the size of data space i.e. 160GB SSD, 120GB data space, 40GB reserved space Fusion I/O has a functionality to change reserved space size # fio-format -s 96G /dev/fct0
Write performance deterioration Write IOPS deterioration (16KB random write) 30000 Continuous write-intensive workloads 25000 20000 IOPS Fastest 15000 Slowest 10000 5000 Stopping writing for a while 0 Intel Fusion(150G) Fusion(120G) Fusion(96G) Fusion(80G) At the beginning, write IOPS was close to “Fastest” line When massive writes happened, write IOPS gradually deteriorated toward “Slowest” line (because massive ERASE happened) Increasing reserved space improves steady-state write throughput Write IOPS recovered to “Fastest” when stopping writes for a long time (Many blocks were ERASEd by background job) Highly depending on Flash memory and I/O controller (TRIM support, ERASE scheduling, etc)
Mitigating write performance deterioration Allocating more reserved space Per space pricing increases. This is not always acceptable Using tachIOn tachIOn is highly optimized for keeping write performance higher Using Software RAID0 Write traffic is distributed to multiple drives. You can gain high enough write throughput even though additional reserved space is not allocated
Sequential I/O Sequential Read/Write throughput (1MB consecutive reads/writes) 600 500 400 MB/s Seq read 300 Seq write 200 100 0 4 HDD(raid10, xfs) Intel(xfs) Fusion(xfs) Typical scenario: Full table scan (read), logging/journaling (write) SSD outperforms HDD for sequential reads, but less significant HDD (4 RAID10) is fast enough for sequential i/o Data transfer size by sequential writes tends to be huge, so you need to care about write deterioration on SSD No strong reason to use SSD for sequential writes
fsync() speed fsync speed 20000 18000 16000 14000 fsync/sec 12000 1KB 10000 8KB 8000 16KB 6000 4000 2000 0 HDD(xfs) Intel (xfs) Fusion I/O(xfs) 10,000+ fsync/sec is fine in most cases Fusion I/O was CPU bound (%system), not I/O bound (%iowait).
HDD is fast for sequential writes / fsync Best Practice: Writes can be boosted by using BBWC/FBWC (Battery Backed up Write Cache), especially for REDO Logs (because it’s sequentially written) No strong reason to use SSDs here seek & rotation time Write cache disk disk seek & rotation time
Filesystem matters Random write iops (16KB Blocks) 20000 18000 16000 14000 12000 1 thread iops 10000 16 thread 8000 6000 4000 2000 0 Fusion(ext3) Fusion (xfs) Fusion (raw) Filesystem On xfs, multiple threads can write to the same file if opened with O_DIRECT, but can not on ext* Good concurrency on xfs, close to raw device ext3 is less optimized for Fusion I/O
Changing I/O unit size Read IOPS and I/O unit size (4 HDD RAID10) 2500 2000 1KB 1500 IOPS 4KB 1000 16KB 500 0 1 2 3 4 5 6 8 10 15 20 30 40 50 100 200 concurrency On HDD, maximum 22% performance difference was found between 1KB and 16KB No big difference when concurrency < 10
Changing I/O unit size on FusionIO Random Read IOPS (FusionIO, SLC) 160000 140000 120000 100000 Reads/s 4KB 80000 8KB 60000 16KB 40000 20000 0 1 2 3 4 5 6 8 10 15 20 30 40 50 100 200 Concurrency Huge difference On SSDs, not only IOPS, but also I/O transfer size matters It’s worth considering that Storage Engines support “configurable block size” functionality
SLC vs MLC (16KB) Random Read IOPS, FusionIO (16KB) 45000 40000 35000 30000 reads/s 25000 SLC 20000 MLC 15000 10000 5000 0 1 2 3 4 5 6 8 10 15 20 30 40 50 100 200 concurrency 8-40% better throughput on SLC
SLC vs MLC (8KB) Random Read IOPS, FusionIO (8KB) 80000 70000 60000 50000 reads/s SLC 40000 MLC 30000 20000 10000 0 1 2 3 4 5 6 8 10 15 20 30 40 50 100 200 concurrency 25-75% better throughput on SLC Random I/O benchmarks should be done carefully (SLC or MLC, I/O unit size, etc..)
tachIOn vs FusionIO (SLC) Random Read IOPS (16KB) 90000 80000 70000 60000 reads/s 50000 FusionIO 40000 tachIOn 30000 20000 10000 0 1 2 3 4 5 6 8 10 15 20 30 40 50 100 200 concurrency tachIOn performs very well, especially when concurrency is high enough
PCI-Express interface and CPU util # cat /proc/interrupts | grep PCI 83: … PCI-MSI vgcinit 202: … PCI-MSI-X eth2-0 210: … PCI-MSI-X eth2-1 218: … PCI-MSI-X eth2-2 226: … PCI-MSI-X eth2-3 234: … PCI-MSI-X eth2-4 # mpstat –P ALL 1 CPU %user %nice %sys %iowait %irq %soft %idle intr/s all 0.45 0.00 7.75 26.69 1.65 0.00 63.45 40046.40 0 1.00 0.00 12.60 86.40 0.00 0.00 0.00 1000.20 1 1.00 0.00 13.63 85.37 0.00 0.00 0.00 0.00 2 0.40 0.00 4.80 26.80 0.00 0.00 68.00 0.00 3 0.00 0.00 0.00 0.00 79.20 0.00 20.80 39033.20 ... %irq (Hardware interrupts) was bottleneck Only one IRQ port was allocated per single PCI-E SSD drive By using multiple drives the number of IRQ ports increases
# of interfaces (tachIOn SLC) Random Read IOPS (16KB, tachIOn) 350000 300000 250000 200000 reads/s Single Drive Two Drives (RAID0) 150000 100000 50000 0 1 2 3 4 5 6 8 10 15 20 30 40 50 100 200 concurrency Two drives result in nearly two times better throughput, if enough read i/o is coming When the number of clients was small (not enough i/o requests were coming), %irq was not so high, so using two drives was not so much helpful The number of slots are limited on most motherboards
# of interfaces (FusionIO MLC) Random Read IOPS (16KB) 70000 60000 50000 reads/s 40000 FusionIO Duo 30000 FusionIO 20000 10000 0 1 2 3 4 5 6 8 10 15 20 30 40 50 100 200 concurrency FusionIO Duo internally has two drives per single PCI-Express connector Two IRQ ports can be used, which greatly increases throughput A couple of restrictions FusionIO Duo has two device files (/dev/fioa, /dev/fiob). Single large native filesystem can not be created FusionIO Duo has physical(height/length) and electrical restrictions – Only one Duo drive can be installed on HP DL360 (two PCI-E ports) physically – On some servers, without an optional power cable, maximum performance can not be gained
tachIOn(SLC) vs FusionIO Duo(MLC) Random Read IOPS (16KB) 90000 80000 70000 60000 tachIOn reads/s 50000 FusionIO Duo 40000 FusionIO 30000 20000 10000 0 1 2 3 4 5 6 8 10 15 20 30 40 50 100 200 concurrency Fusion IO Duo performs good enough, considering it’s MLC
Opteron vs Nehalem(tachIOn) tachIOn Random Read IOPS (16KB) 90000 80000 70000 60000 reads/s 50000 Nehalem X5650 40000 Opteron 6174 30000 20000 10000 0 1 2 3 4 5 6 8 10 15 20 30 40 50 100 200 # of threads 2 times performance difference Single thread IOPS is bad Latency difference between PCI-E and CPU seems quite high
Opteron vs Nehalem(tachIOn) tachIOn Random Read IOPS(4KB) 300000 250000 200000 reads/s Nehalem X5650 150000 Opteron 6174 100000 50000 0 1 2 3 4 5 6 8 10 15 20 30 40 50 100 200 # of threads 2-4 times performance difference In both cases %irq was 100% used Handling Hardware Interrupts requires high CPU clock speed
How about using SSD as L2 Cache? Using SSD as L2Cache Interesting approach, as long as SSD is costly and capacity is limited Transparent from applications Recent lineups ZFS L2ARC – Part of ZFS filesystem Facebook FlashCache – Working as Linux Kernel module FusionIO DirectCache – Working between OS and FusionIO. Depending on FusionIO drives Oracle Smart Flash Cache – L2 cache of Oracle Database. Depending on Oracle database Issues Performance is not good (FlashCache) – Overheads on Linux Kernel seem huge – Even though data is 100% on SSD, random read iops dropped 40%, random write iops dropped 75% on FusionIO (tested with CentOS5.5) – Less performance drops on Intel X25-E In practice it’s Single Point of Failure – It’s not just a cache. We expect SSD-level performance. If it’s broken, the system will go down. Total write volume grows significantly. L2Cache needs to be written when reading from HDD
Virtualization? Currently performance drops are serious (on faster drives) Got only 1/30 throughput when tested with Ubuntu 10.4 + KVM + FusionIO When tested with HDD, performance drops were not so serious Not many case studies about running FusionIO/tachIOn on virtualization environments It’s better to run multiple MySQL instances on single OS
Virtualization benchmarks (HDD) Random Read IOPS (Dell could HDD, 16KB unit) 5000 4500 4000 3500 reads/sec 3000 Physical 2500 2000 KVM 1500 1000 500 0 1 2 3 4 5 6 8 10 15 20 30 40 50 100 200 # of threads Not so bad on HDD
Virtualization benchmarks (SATA SSD) Random Read IOPS (Dell cloud SSD, 16KB unit) 25000 20000 reads/sec 15000 Physical KVM 10000 5000 0 1 2 3 4 5 6 8 10 15 20 30 40 50 100 200 # of threads Difference becomes huge when using SSD
MySQL Deployment Practices on SSD What will happen if Replacing HDD with Intel SSD (SATA) Replacing HDD with Fusion I/O (PCI-E) Moving log files and ibdata to HDD Not using Nehalem MySQL 5.5 or 5.1 DBT-2 benchmarks (write intensive) 200 Warehouses (20GB – 25GB hot data) Fusion I/O: SLC, 96GB data space, 64GB reserved space
HDD vs Intel SSD vs Fusion I/O HDD Intel Fusion I/O Buffer pool 1G 1125.44 5709.06 15122.75 NOTPM: Number of Transactions per minute Fusion I/O is a PCI-E based SSD PCI-E is much faster than SAS/SATA 14x improvement compared to 4HDDs
Which should we spend money, RAM or SSD? HDD Intel Fusion I/O Buffer pool 1G 1125.44 5709.06 15122.75 Buffer pool 2G 1863.19 Buffer pool 5G 4385.18 Buffer pool 30G 36784.76 (Caching all hot data) Increasing RAM (buffer pool size) reduces random disk reads Because more data are cached in the buffer pool If all data are cached, only disk writes (both random and sequential) happen Disk writes happen asynchronously, so application queries can be much faster Large enough RAM + HDD outperforms too small RAM + SSD
Which should we spend money, RAM or SSD? HDD Intel Fusion I/O Buffer pool 1G 1125.44 5709.06 15122.75 Buffer pool 2G 1863.19 7536.55 20096.33 Buffer pool 5G 4385.18 12892.56 30846.34 Buffer pool 30G 36784.76 - 57441.64 (Caching all hot data) It is not always possible to cache all hot data Fusion I/O + good amount of memory (5GB) was pretty good Basic rule can be: If you can cache all active data, large enough RAM + HDD If you can’t, or if you need extremely high throughput, spend on both RAM and SSD
MySQL file location SSD is extremely good at random reads SSD is very good at random writes HDD is good enough at sequential reads/writes No strong reason to use SSD for sequential writes Random I/O oriented: Data Files (*.ibd) – Sequential reads if doing full table scan Undo Log, Insert Buffer (ibdata) – UNDO tablespace (small in most cases, except for running long-running batch) – On-disk insert buffer space (small in most cases, except that InnoDB can not catch up with updating indexes) Sequential Write oriented: Doublewrite Buffer (ibdata0) – Write volume is equal to *ibd files. Huge Binary log (mysql-bin.XXXXXX) Redo log (ib_logfile) Backup files
Moving sequentially written files into HDD Fusion I/O Fusion I/O + HDD Up Buffer pool 15122.75 19295.94 +28% 1G (us=25%, wa=15%) (us=32%, wa=10%) Buffer pool 20096.33 25627.49 +28% 2G (us=30%, wa=12.5%) (us=36%, wa=8%) Buffer pool 30846.34 39435.25 +28% 5G (us=39%, wa=10%) (us=49%, wa=6%) Buffer pool 57441.64 66053.68 +15% 30G (us=70%, wa=3.5%) (us=77%, wa=1%) Moving ibdata, ib_logfile, (+binary logs) into HDD High impact on performance Write volume to SSD becomes half because doublewrite area is allocated in HDD %iowait was significantly reduced Write volume on SSD is less than half
Does CPU matter? Nehalem Older Xeon CPUs Memory CPUs QPI: 25.6GB/s FSB: 10.6GB/s North Bridge North Bridge (IOH) Memory (MCH) PCI-Express PCI-Express Nehalem has two big advantages 1. Memory is directly attached to CPU : Faster for in-memory workloads 2. Interface speed between CPU and North Bridge is 2.5x higher, and interface traffics do not conflict with CPU<->Memory workloads Faster for disk i/o workloads when using PCI-Express SSDs
Harpertown X5470 (older Xeon) vs Nehalem X5570 (HDD) HDD Harpertown X5470, Nehalem(X5570, Up 3.33GHz 2.93GHz) Buffer pool 1G 1135.37 (us=1%) 1125.44 (us=1%) -1% Buffer pool 2G 1922.23 (us=2%) 1863.19 (us=2%) -3% Buffer pool 5G 4176.51 (us=7%) 4385.18(us=7%) +5% Buffer pool 30G 30903.4 (us=40%) 36784.76 (us=40%) +19% us: userland CPU utilization CPU difference matters on CPU bound workloads
Harpertown X5470 vs Nehalem X5570 (Fusion) Fusion I/O+HDD Harportown X5470, Nehalem(X5570, Up 3.33GHz 2.93GHz) Buffer pool 1G 13534.06 (user=35%) 19295.94 (user=32%) +43% Buffer pool 2G 19026.64 (user=40%) 25627.49 (user=37%) +35% Buffer pool 5G 30058.48 (user=50%) 39435.25 (user=50%) +31% Buffer pool 30G 52582.71 (user=76%) 66053.68 (user=76%) +26% TPM difference was much higher than HDD For disk i/o bound workloads (buffer pool 1G/2G), CPU utilizations on Nehalem were smaller, but TPM were much higher Verified that Nehalem is much more efficient for PCI-E workloads Benefit from high interface speed between CPU and PCI-Express Fusion I/O fits with Nehalem much better than with traditional CPUs
Intel SSDs with a traditional H/W raid controller Single raw Intel Four RAID5 Intel Down Buffer pool 1G 5709.06 2975.04 -48% Buffer pool 2G 7536.55 4763.60 -37% Buffer pool 5G 12892.56 11739.27 -9% Raw SSD drives performed much better than using a traditional H/W raid controller Even on RAID10 performance was worse than single raw drive H/W Raid controller seemed serious bottleneck Make sure SSD drives have write cache and capacitor itself (Intel X25- V/M/E doesn’t have capacitor) Use JBOD + write cache + capacitor Intel 320 SSD supports capacitor
Enable HyperThreading Fusion I/O + HDD HT OFF (8) HT ON (16) Up Buffer pool 1G 19295.94 20785.42 +7.7% Buffer pool 2G 25627.49 28438.00 +11% Buffer pool 5G 39435.25 45785.12 +16% Buffer pool 30G 66053.68 81412.23 +23% InnoDB Plugin and 5.1 scales well with 16-24 CPU cores HT is more effective on SSD environments because loads are more CPU bound
MySQL 5.5 Fusion I/O + HDD MySQL5.1 MySQL5.5 Up Buffer pool 1G 19295.94 24019.32 +24% Buffer pool 2G 25627.49 32325.76 +26% Buffer pool 5G 39435.25 47296.12 +20 Buffer pool 30G 66053.68 67253.45 +1.8% Got 20-26% improvements for disk i/o bound workloads on Fusion I/O Both CPU %user and %iowait were improved – %user: 36% (5.5.2) to 44% (5.5.4) when buf pool = 2g – %iowait: 8% (5.5.2) to 5.5% (5.5.4) when buf pool = 2g, but iops was 20% higher Could handle a lot more concurrent i/o requests in 5.5 ! No big difference was found on 4 HDDs – Works very well on faster storages such as Fusion I/O, lots of disks
Where can PCI-Express SSD be used? Deploying on master? If PCI-E SSD is used on master, replication delay will happen again – 10,000IOPS from single thread, 40,000+ IOPS from 100 threads 10,000IOPS from 100 threads can be achieved from SATA SSD Parallel SQL threads should be implemented in MySQL Deploying on slave? If using HDD on master, SATA SSD should be enough to handle workloads – PCI-Express SSD is much more expensive than SATA SSD How about running multiple MySQL instances on single server? – Virtualization is not fast – Running multiple MySQL instances on single OS is more reasonable Does PCI-E SSD have enough storage capacity to run multiple instances? On HDD environments, typically only 100-200GB of database data can be stored because of slow random IOPS on HDD FusionIO SLC: 320GB Duo + 160GB = 480GB FusionIO MLC: 1280GB Duo + 640GB = 1920GB (or using ioDrive Octal) tachIOn SLC: 800GB x 2 = 1600GB
Running multiple slaves on single boxBefore After M M B M M B B S1 S2 S3 B S1 S2 S3 S1, S1 S2, S2 S1, S1 S2, S2 M M B S1 S2 S3 B S1 S2 S3 B M M B Running multiple slaves on a single PCI-E slave Master and Backup Server are still HDD based Consolidating multiple slaves Since slave’s SQL thread is single threaded, you can gain better concurrency by running multiple instances The number of instances is mainly restricted by capacity
Our environment Machine HP DL360G7 (1U), or Dell R610 PCI-E SSD FusionIO MLC (640-1280GB Duo + 320-640GB ) tachIOn SLC (800GB x 2) -> MLC CPU Two sockets, Nehalem 6-core per socket, HT enabled – 24 logical CPU cores are visible – Four socket machine is too expensive RAM 60GB or more Network Broadcom BCM5709, Four ports Bonding x 2 Two physical IP addresses, 6-12 (=instances) virtual IP addresses HDD 4-8 SAS RAID1+0 For backups, redo logs, relay logs, (optionally) doublewrite buffer
Statistics Consolidating 7 instances on FusionIO (640GB MLC Duo + 320GB MLC) Let half of SELECT queries go to these slaves 6GB innodb_buffer_pool_size Peak QPS (total of 7 instances) 61683.7 query/s 37939.1 select/s 7861.1 update/s 1105 insert/s 1843 delete/s 3143.5 begin/s CPU Utilization %user 27.3%, %sys 11%(%soft 4%), %iowait 4% C.f. SATA SSD：%user 4%, %sys 1%, %iowait 1% Buffer pool hit ratio 99.4% SATA SSD : 99.8% No replication delay No significant (100+ms) response time delay caused by SSD
CPU loads 22:10:57 CPU %user %nice %sys %iowait %irq %soft %steal %idle intr/s 22:11:57 all 27.13 0.00 6.58 4.06 0.14 3.70 0.00 58.40 56589.95 … 22:11:57 23 30.85 0.00 7.43 0.90 1.65 49.78 0.00 9.38 44031.82 CPU utilization was high, but should be able to handle more %user 27.3%, %sys 11%(%soft 4%), %iowait 4% Reached storage capacity limit (960GB). Using 1920GB MLC should be fine to handle more instances Network will be the first bottleneck Recv: 14.6MB/s, Send: 28.7MB/s CentOS5 + bonding is not good for network requests handling (only single CPU core can handle requests) Using four or more network cables, then building two or more bond interfaces should scale
Things to consider Allocating different IP addresses or port numbers Administration tools are also affected We allocated different IP addresses because some of existing tools depend on “port=3306” ip addr bind-address=“virtual ip address” in my.cnf Creating separated directories and files Socket files, data directories, InnoDB files, binary log files etc should be stored on different location each other Storing some files on HDD, others on SSD Binary logs, Relay logs, Redo logs, error/slow logs, ibdata0 (files where doublewrite buffer is written), backup files on HDD Others on SSD
taskset # taskset -pc 0,12,2,14 `cat /var/lib/mysql/mysqld1.pid` # taskset -pc 1,13,3,15 `cat /var/lib/mysql/mysqld2.pid` # taskset -pc 5,17,7,19 `cat /var/lib/mysql/mysqld3.pid` # taskset -pc 8,20,10,22 `cat /var/lib/mysql/mysqld4.pid` # taskset -pc 9,21,11,23 `cat /var/lib/mysql/mysqld5.pid` # taskset -pc 4,16,6,18 `cat /var/lib/mysql/mysqld6.pid` # taskset -pc 8,20,10,22 `cat /var/lib/mysql/mysqld7.pid` MySQL server currently does not scale well with 24 logical CPU cores When running 6+ instances on 24 cores, the number of utilized CPU cores should be limited per instance Check /proc/cpuinfo output Use same physical id (socket id) for the same instance Within the same physical id, use same core id (physical core) for the same instance
Application Design Bottleneck shifts from storage to CPU/Network Massively executed SQL statements should be migrated to NoSQL / HandlerSocket /etc Separating tables History tables (only recent data is accessed) fit with HDD because most of active data fit in memory Rarely accessed tables can also be stored on HDD Other tables on SSD
Making MySQL better Parallel SQL threads Pool of threads 8KB/4KB InnoDB Blocks Minimized performance stalls No hot global/large mutex LINEAR HASH partitions on large tables helps – Index mutex is allocated per partition
Future improvements from DBA perspective One master, one backup/slave and one DR slave Single slave should be enough from performance perspective Differential recovery (no full restore) on crash Restoring 800GB data file takes hours in total MySQL should be better for crashes Master should flush transactional data to disks at commit – sync-binlog=1, innodb_flush_log_at_trx_commit=1 – Right now performance is poor Semi-sync replication (GA in 5.5) Crash-safe binary logs (binlog checksum) Relay log execution status should be stored on InnoDB tables, not on file(relay-log.info) (in 5.6?) – Relay log status can be consistent
Conclusion for choosing H/W Disks PCI-E SSDs (i.e. FusionIO, tachIOn) perform very well SAS/SATA SSDs with capacitor (i.e. Intel 320) Carefully research RAID controller. Many controllers do not scale with SSD drives Keep enough reserved space, use tachIOn, or use RAID0 if you need to handle massive write traffics HDD is good at sequential writes Use Nahalem CPU Especially when using PCI-Express SSDs
Conclusion for database deployments Put sequentially written files on HDD ibdata, ib_logfile, binary log files HDD is fast enough for sequential writes Write performance deterioration can be mitigated Life expectancy of SSD will be longer Put randomly accessed files on SSD *ibd files, index files(MYI), data files(MYD) SSD is 10x -100x faster for random reads than HDD Archive less active tables/records to HDD SSD is still much expensive than HDD Use 5.1 InnoDB Plugin or 5.5 Higher scalability & concurrency matters on faster storage
What will happen in the real database world? These are just my thoughts.. Less demand for NoSQL Isn’t it enough for many applications just to replace HDD with Fusion I/O? Importance on functionality will be relatively stronger Stronger demand for Virtualization Single server will have enough capacity to run two or more mysqld instances Right now performance is poor. Running multiple instances is a good workaround I/O volume matters Not just IOPS Configurable block size, disabling doublewrite, etc Concurrency matters Single SSD scales as well as 8-16 HDDs Concurrent ALTER TABLE, parallel query
Memory and Swap Space Management
Random Access Memory RAM access speed is much faster than HDD/SSD RAM: -60ns – 100,000 queries per second is not impossible HDD: -5ms SSD: 100-500us 16-100+GB RAM is now pretty common *hot application data* should be cached in memory Minimizing hot application data size is important Use compact data types (SMALLINT instead of VARCHAR/BIGINT, TIMESTAMP instead of DATETIME, etc) Do not create unnecessary indexes Delete records or move to archived tables, to keep hot tables smaller
Cache hot application data in memory DBT-2 (W200) Transactions per Minute %user %iowait Buffer pool 1G 1125.44 2% 30% Buffer pool 2G 1863.19 3% 28% Buffer pool 5G 4385.18 5.5% 33% Buffer pool 30G 36784.76 36% 8% (All data in cache) DBT-2 benchmark (write intensive) 20-25GB hot data (200 warehouses, running 1 hour) RAM size affects everything. Not only for SELECT, but also for INSERT/UPDATE/DELETE INSERT: Random reads/writes happen when inserting into indexes in random order UPDATE/DELETE: Random reads/writes happen when modifying records
Use Direct I/O Buffered I/O Direct I/O InnoDB Buffer Pool InnoDB Buffer Pool Filesystem Cache RAM RAM InnoDB Data File InnoDB Data File Direct I/O is important to fully utilize Memory innodb_flush_method=O_DIRECT Alignment: File i/o unit must be a factor of 512 bytes Can’t use O_DIRECT for InnoDB Log File, Binary Log File, MyISAM, PostgreSQL data files, etc
Do not allocate too much memory user$ top Mem: 32967008k total, 32808696k used, 158312k free, 10240k buffers Swap: 35650896k total, 4749460k used, 30901436k free, 819840k cached PID USER PR NI VIRT RES SHR S %CPU %MEM TIME+ COMMAND 5231 mysql 25 0 35.0g 30g 324 S 0.0 71.8 7:46.50 mysqld What happens if no free memory space is available? Reducing filesystem cache to allocate memory space Swapping process(es) to allocate memory space Swap is bad Process spaces are written to disk (swap out) Disk reads happen when accessing on-disk process spaces (swap in) Massive random disk reads and writes will happen
What if setting swap size to zero? By setting swap size to zero, swap doesn’t happen anymore. But.. Very dangerous When neither RAM nor swap space is available, OOM killer is invoked. OOM Killer may kill any process to allocate memory space The most memory-consuming process (mysqld) will be killed at first It’s abort shutdown. Crash recovery takes place at restart Priority is determined by ORDER BY /proc/<PID>/oom_score DESC Normally mysqld has the highest score – Depending on VMsize, CPU time, running time, etc It often takes very long time (minutes to hours) for OOM Killer to kill processes We can’t do anything until enough memory space is available
Do not set swap=zero top - 01:01:29 up 5:53, 3 users, load average: 0.66, 0.17, 0.06 zombie Tasks: 170 total, 3 running, 167 sleeping, 0 stopped, 0 zombie Cpu(s): 24.9%sy, Cpu(s): 0.0%us, 24.9%sy, 0.0%ni,75.0%id,0.2%wa,0.0%hi, 0.0%si,0.0%st Mem: Mem: 32967008k total, 32815800k used, 151208k free, buffers 8448k buffers Swap: 0k total, 0k used, 0k free, 376880k cached PID USER PR NI VIRT RES SHR S %CPU %MEM COMMAND TIME+ COMMAND 26988 mysql 25 0 30g 30g 1452 R 98.5 97.7 0:42.18 mysqld If no memory space is available, OOM killer will be invoked Some CPU cores consume 100% system resources 24.9% (average) = 1 / 4 core use 100% cpu resource in this case Terminal freezed (SSH connections can’t be established) Swap is bad, but OOM killer is much worse than swap
What if stopping OOM Killer? If /proc/<PID>/oom_adj is set to -17, OOM Killer won’t kill the process Setting -17 to sshd is a good practice so that we can continue remote login # echo -17 > /proc/<pid of sshd>/oom_adj But don’t set -17 to mysqld If over-memory-consuming process is not killed, Linux can’t have any available memory space We can’t do anything for a long long time.. -> Long downtime
Swap space management Swap space is needed to stabilize systems But we don’t want mysqld swapped out What consumes memory? RDBMS – Mainly process space is used (innodb_buffer_pool, key_buffer, sort_buffer, etc) – Sometimes filesystem cache is used (InnoDB Log Files, binary/relay log files, MyISAM files, etc) Administration (backup, etc) – Mainly filesystem cache is used We want to keep mysqld in RAM, rather than allocating large filesystem cache
Be careful about backup operations Mem: 32967008k total, 28947472k used, 4019536k free, 152520k buffers Swap: 35650896k total, 0k used, 35650896k free, 197824k cached PID USER PR NI VIRT RES SHR S %CPU %MEM TIME+ COMMAND 5231 mysql 25 0 27.0g 27g 288 S 0.0 92.6 7:40.88 mysqld Copying 8GB datafile Mem: 32967008k total, 32808696k used, 158312k free, 10240k buffers Swap: 35650896k total, 4749460k used, 30901436k free, 8819840k cached PID USER PR NI VIRT RES SHR S %CPU %MEM TIME+ COMMAND 5231 mysql 25 0 27.0g 22g 324 S 0.0 71.8 7:46.50 mysqld Copying large files often causes swap
vm.swappiness = 0 Mem: 32967008k total, 28947472k used, 4019536k free, 152520k buffers Swap: 35650896k total, 0k used, 35650896k free, 197824k cached PID USER PR NI VIRT RES SHR S %CPU %MEM TIME+ COMMAND 5231 mysql 25 0 27.0g 27g 288 S 0.0 91.3 7:55.88 mysqld Copying 8GB of datafile Mem: 32967008k total, 32783668k used, 183340k free, 3940k buffers Swap: 35650896k total, 216k used, 35650680k free, 4117432k cached PID USER PR NI VIRT RES SHR S %CPU %MEM TIME+ COMMAND 5231 mysql 25 0 27.0g 27g 288 S 0.0 80.6 8:01.44 mysqld Set vm.swappiness=0 in /etc/sysctl.conf Default is 60 When physical RAM was fully consumed, Linux kernel reduces filesystem cache with high priority (lower swappiness increases priority) After no file system cache is available, swapping starts OOM killer won’t be invoked if large enough swap space is allocated. It’s safer
But swap happens even though swappiness==0 32GB RAM box, CentOS5 (2.6.18.128) innodb_buffer_pool_size = 26GB mysqld other than innnodb_buf_pool_size: approx 3GB Kernel space : approx 1GB top - 11:54:51 up 7 days, 15:17, 1 user, load average: 0.21, 0.14, 0.10 Tasks: 251 total, 1 running, 250 sleeping, 0 stopped, 0 zombie Cpu(s): 0.5%us, 0.2%sy, 0.0%ni, 98.9%id, 0.3%wa, 0.0%hi, 0.1%si, 0.0%st Mem: 32818368k total, 31154696k used, 1663672k free, 125048k buffers Swap: 4184924k total, 1292756k used, 2892168k free, 2716380k cached PID USER PR NI VIRT RES SHR S %CPU %MEM TIME+ COMMAND 6999 mysql 15 0 28.4g 26g 5444 S 9.3 85.2 210:52.67 mysqld
Swap caused by log files InnoDB Buffer Pool Filesystem Cache InnoDB Data File Log Files Swap happened even though filesystem cache was not wiped out InnoDB log file size : 3GB in total It’s common in InnoDB Plugin because crash recovery becomes much faster Binary logs (or relay logs)
Swap bug on Linux A known issue of Linux Kernel Fixed in 2.6.28: Merged on RHEL6 https://bugzilla.redhat.com/show_bug.cgi?id=160033 When will CentOS 6 be released? Before the bug was fixed: swap_tendency = mapped_ratio / 2 + distress + sc->swappiness; If swap_tendency > 100, swap happens regardless of filesystem cache – This might happen if distress is high enough, even though swappiness == 0 When distress becomes high? Writing to InnoDB log files, binary log files, relay log files On recent MySQL this issue becomes even more serious because innodb_log_file_size can be bigger – Slow crash recovery bug has been fixed in InnoDB Plugin, so using 3-4GB inodb log file becomes common What workarounds can be available in RHEL5? Decrease innodb_buffer_pool_size or innodb_log_file_size? – I don’t like the solution Unmap InnoDB log files and binary/relay log files from file system cache
Tool: unmap_mysql_logs https://github.com/yoshinorim/unmap_mysql_logs Perl based command line tool – Depending on Inline::C CPAN module Unmapping the below files from filesystem cache InnoDB Log Files: All files Binary log files, Relay log files: – Current (hot) binlog/relay log: All except the latest (tail) 10% of the file – Because the tail of the current files are read by Binlog Dump thread or SQL thread: needs to be cached – Others: all InnoDB Buffer Pool Filesystem Cache InnoDB Data File Log Files
Performance effect Mem: 32825116k total, 32728924k used, 96192k free, 246360k buffers Swap: 4186072k total, 118764k used, 4067308k free, 2803088k cached PID USER PR NI VIRT RES SHR S %CPU %MEM TIME+ COMMAND 30814 mysql 15 0 28.8g 27g 5644 S 81.1 89.4 158198:37 mysqld Mem: 32825116k total, 30061632k used, 2763484k free, 247420k buffers Swap: 4186072k total, 118764k used, 4067308k free, 143372k cached PID USER PR NI VIRT RES SHR S %CPU %MEM TIME+ COMMAND 30814 mysql 15 0 28.8g 27g 5644 S 76.1 89.4 158200:59 mysqld Filesystem cache was reduced, free memory space was increased No swapping has happened anymore Regularly invoke the script Every 10 minutes from cron
Memory allocator mysqld uses malloc()/mmap() for memory allocation Faster and more concurrent memory allocator such as tcmalloc can be used Install Google Perftools (tcmalloc is included) – # yum install libunwind – # cd google-perftools-x.y ; ./configure --enable-frame-pointers; make; make install export LD_PRELOAD=/usr/local/lib/tcmalloc_minimal.so; mysqld_safe & InnoDB internally uses its own memory allocator Can be changed in InnoDB Plugin – If Innodb_use_sys_malloc = 1(default 1), InnoDB uses OS memory allocator – tcmalloc can be used by setting LD_PRELOAD
Memory allocator would matter for CPU bound workloads Default tcmalloc_minimal %user up allocator Buffer pool 1G 1125.44 1131.04 2% +0.50% Buffer pool 2G 1863.19 1881.47 3% +0.98% Buffer pool 5G 4385.18 4460.50 5.5% +1.2% Buffer pool 30G 36784.76 38400.30 36% +4.4% DBT-2 benchmark (write intensive) 20-25GB hot data (200 warehouses, running 1 hour)
Be careful about per-session memory Do not allocate much more memory than needed (especially for per-session memory) Allocating 2MB takes much longer time than allocating 128KB Linux malloc() internally calls brk() if size <= 512KB, else calling mmap() In some cases too high per-session memory allocation causes negative performance impacts SELECT * FROM huge_myisam_table LIMIT 1; SET read_buffer_size = 256*1024; (256KB) – -> 0.68 second to run 10,000 times SET read_buffer_size = 2048*1024; (2MB) – -> 18.81 seconds to run 10,000 times In many cases MySQL does not allocate per-session memory than needed. But be careful about some extreme cases (like above: MyISAM+LIMIT+FullScan)
File I/O
File I/O and synchronous writes RDBMS calls fsync() many times (per transaction commit, checkpoints, etc) Make sure to use Battery Backed up Write Cache (BBWC) on raid cards 10,000+ fsync() per second, without BBWC less than 200 on HDD Disable write cache on disks for safety reasons Do not set “write barrier” on filesystems (enabled by default in some cases) Write-through to disks even though BBWC is enabled (very slow) ext3: mount -o barrier=0 (needed for some distros such as SuSe Linux) xfs: mount -o nobarrier drbd: no-disk-barrier in drbd.conf seek & rotation time Write cache with battery disk disk seek & rotation time
BBWC and “Auto Learn” Issues Battery needs to be discharged and recharged every specific periods “Auto Learn” Disabling write cache (entering write through), recharging battery – Takes 1 Hour or even more Sudden performance drops happen – iowait suddenly increases a lot Automated auto learn – Default behavior – It is serious if happened at peak time Manual auto learn (in many cases officially not supported) – Manually discharge and recharge on off-peak loads – If you forget to execute for a long time (i.e. 8 months), battery will be broken Expected behavior on Dell servers (not a bug) LSI RAID Controllers with BBWC Approximately every 90 days – “Event Description: BBU disabled; changing WB virtual disks to WT”
FBWC Flash Back Write Cache Using NOR flash memory and capacitor instead of battery No “Auto Learn” issue HP adopts FBWC on recent lineups (DL360G7)
Overwriting or Appending? Some files are overwritten (fixed file size), others are appended (increasing file size) Overwritten: InnoDB Logfile Appended: Binary Logfile Appending + fsync() is much slower than overwriting + fsync() – Additional file space needs to be allocated & file metadata needs to be flushed per fsync() 10,000+ fsync/sec for overwriting, 3,000 or less fsync/sec for appending – Appending speed highly depends on filesystems – Copy-on-write filesystems such as Solaris ZFS is fast enough for appending (7,000+) Be careful when using sync-binlog=1 for binary logs – Consider using ZFS – Check “preallocating binlog” worklog: WL#4925 Do not extend files too frequently – innodb-autoextend-increment = 20 (default 8)
Quick file i/o health check Checking FBWC/BBWC is enabled, and write barrier is disabled Overwriting + fsync() test – Run mysqlslap insert(InnoDB, single-threaded, innodb_flush_log_at_trx_commit=1), check qps is over 1,000 $ mysqlslap --concurrency=1 --iterations=1 --engine=innodb ¥ --auto-generate-sql --auto-generate-sql-load-type=write ¥ --number-of-queries=100000
Buffered and asynchronous writes Some file i/o operations are not direct i/o, not synchronous file copy, MyISAM, mysqldump, innodb_flush_log_at_trx_commit=2, etc Dirty pages in filesystem cache needs to be flushed to disks in the end pdflush takes care of it, maximum 8 threads When? -> highly depending on vm.dirty_background_ratio and vm.dirty_ratio Flushing dirty pages starts in background after reaching dirty_background_ratio * RAM (Default is 10%, 10% of 64GB is 6.4GB) Forced flush starts after reaching dirty_ratio * RAM (Default is 40%) Forced, and burst dirty page flushing is problematic All buffered write operations become synchronous, which hugely increase latency Do flush dirty pages aggressively Execute sync; while doing massive write operations Reduce vm.dirty_background_ratio Upgrade to 2.6.32 or higher – pdflush threads are allocated per device. Flushing to slow devices won’t block other pdflush threads
Filesystem – ext3 By far the most widely used filesystem But not always the best Deleting large files takes long time Internally has to do a lot of random disk i/o (slow on HDD) In MySQL, if it takes long time to DROP table, all client threads will be blocked to open/close tables (by LOCK_open mutex) Be careful when using MyISAM, InnoDB with innodb_file_per_table, PBXT, etc Writing to a file is serialized Serialized by “i-mutex”, allocated per i-node Sometimes it is faster to allocate many files instead of single huge file Less optimized for faster storage (like PCI-Express SSD) Use “dir_index” to speed up searching files Use barrier=0 to disable write-through
Tip: fast large file remove Creating hard links ln /data/mysql/db1/huge_table.ibd /var/tmp/huge_table.ibd.link DROP TABLE huge_table; rm –f /var/tmp/huge_table.ibd.link Physical file remove doesn’t happen because ref count is not zero When removing the hard link, physical file remove happens because at that time ref count is zero Physical remove operation causes massive random disk i/o, but at this stage this doesn’t take MySQL mutex so it won’t block MySQL operations – But iowait increases so don’t run rm command when load is high
Filesystem – xfs/ext2 xfs Fast for dropping files Concurrent writes to a file is possible when using O_DIRECT Not officially supported in current RHEL (Supported in SuSE) Disable write barrier by setting “nobarrier” ext2 Faster for writes because ext2 doesn’t support journaling It takes very long time for fsck On active-active redundancy environment (i.e. MySQL Replication), in some cases ext2 is used to gain performance Btrfs (under development) Copy-on-write filesystem Supporting transactions (no half-block updates) Snapshot backup with no overhead
Concurrent write matters on fast storage 1 i/o thread Random Write IOPS (16KB Blocks) 100 i/o threads 20000 18000 16000 14000 12000 10000 8000 6000 4000 2000 0 HDD(ext3) HDD(xfs) Intel(ext3) Intel(xfs) Fusion(ext3) Fusion (xfs) Negligible on HDD (4 SAS RAID1) 1.8 times difference on Fusion I/O
I/O scheduler Note: RDBMS (especially InnoDB) also schedules I/O requests so theoretically Linux I/O scheduler is not needed Linux has I/O schedulers to efficiently handle lots of I/O requests “I/O scheduler type” and “Queue Size” matters Types of I/O schedulers (introduced in 2.6.10: RHEL5) noop: Sorting incoming i/o requests by logical block address, that’s all deadlilne: Prioritize read (sync) requests rather than write requests (async) to some extent (to avoid “write-starving-reads” problem) cfq(default): Fairly scheduling i/o requests per i/o thread anticipatory: Removed in 2.6.33 (bad scheduler. Don’t use it) Default is cfq, but noop / deadline is better in many cases # echo noop > /sys/block/sdX/queue/scheduler
cfq madness Running two benchmark programs concurrently 1. Multi-threaded random disk reads (Simulating RDBMS reads) 2. Single-threaded overwriting + fsync() (Simulating redo log writes) Random Read write+fsync( Scheduler reads/sec from writes/sec from i/o threads ) running iostat iostat 1 No noop/deadline 260 0 cfq 260 0 100 No noop/deadline 2100 0 cfq 2100 0 1 Yes noop/deadline 212 14480 cfq 248 246 100 Yes noop/deadline 1915 12084 cfq 2084 0 In RDBMS, write IOPS is often very high because HDD + write cache can handle thousands of transaction commits per second (write+fsync) Write iops was adjusted to per-thread read iops in cfq, which reduced total iops significantly
Changing I/O scheduler (InnoDB) DBT-2 (MySQL5.1) 15000 RAID1+0 10000 NOTPM RAID5 5000 0 noop cfq deadline as - Sun Fire X4150 (4 HDDs, H/W RAID controller+BBWC) - RHEL5.3 (2.6.18-128) - Built-in InnoDB 5.1
Changing I/O scheduler queue size (MyISAM) Time to insert 1 million records (HDD) 5000 4000 Seconds 3000 queue size=100000 2000 queie size=128 (default) 1000 0 1 5 9 13 17 21 25 29 33 37 41 45 49 53 57 61 65 69 73 77 Existing records (millions) Queue size = N Sorting N outstanding I/O requests to optimize disk seeks MyISAM does not optimize I/O requests internally Highly depending on OS and storage When inserting into indexes, massive random disk writes/reads happen Increasing I/O queue size reduces disk seek overheads # echo 100000 > /sys/block/sdX/queue/nr_requests No impact in InnoDB Many RDBMS including InnoDB internally sort I/O requests
Network
Fighting against network bottlenecks Latency 100Mbit Ethernet: 3000 us – Around 35,000qps from 100 memcached clients – Not good if you use SSD – Easily reaches 100Mbps bandwidth when copying large files 1Gbit Ethernet: 400us – Around 250,000qps from 100 memcached clients – 100,000qps is not impossible in MySQL Latency is not so amazing at 10Gbit Ethernet – Check Dolphin Supersockets, etc Single CPU core might be bottleneck Older Linux kernels (including CentOS5) do not scale with CPU cores for network requests handling – Use latest kernels (RHEL6 etc) when this becomes really issue
Tx/Rx multiqueue # cat /proc/interrupts | grep PCI 202: … PCI-MSI-X eth2-0 210: … PCI-MSI-X eth2-1 218: … PCI-MSI-X eth2-2 226: … PCI-MSI-X eth2-3 234: … PCI-MSI-X eth2-4 Network requests can be distributed between CPU cores Some NICs and drivers work multiqueue on older kernels, but don’t work when bonding is enabled Check /proc/interrupts Check %irq and %soft from mpstat Single CPU core might occupy 100% %soft Use latest kernels (RHEL6 etc) when this becomes really issue
Be careful about bandwidth Between servers, racks, switches Be careful when sending large data Setting up OS from PXE server Sending large backup file Restoring large data and initializing replication Do not use 100Mbps networks Network switch, Network Cards, Cables When setting up new replication slave, the master sends large binlog events to slaves, which easily occupies 100Mbps traffics START/STOP SLAVE IO_THREADS; repeatedly Manually send binlogs by scp with upper limits, then apply events by mysqlbinlog and mysql
Remote datacenter RTT (round trip time) from Tokyo to West coast easily exceeds 200 ms MySQL Replication for Disaster Recovery purpose I/O thread delays RTT/2 – Time to send binary log from master to slave – Not bad (ping delays RTT) – Lots of binary log events can be sent at once Do not use semi-synchronous replication here – Every commit takes RTT for send/recv. Only few transactions can be handled per second
Waiting three seconds for SYN Establishing network connections is expensive Client: Sending TCP SYN Server: Sending TCP SYN+ACK Client: Sending TCP ACK If client fails to receive SYN+ACK packet, it waits for three seconds, then sending SYN again This happens quite a lot of times on high traffic environments Not limited to MySQL Increasing backlog helps, but problems still exist – echo 8192 > /proc/sys/net/ipv4/tcp_max_syn_backlog – back_log in my.cnf Using persistent connections / connection pooling is the fundamental solution
Flows to establish N/W connections Client Server1. Sending SYN,Changing state to SYN_SENT 2. Receiving SYN 3. Checking conditions (i.e. back_log) If it doesn’t meet criteria, dropping it 4. Generating SYN+ACK, Changing state to SYN_RECV 5. Sending SYN+ACK 6. Receiving SYN+ACK 7. Generating ACK 8. Sending ACK 9. Receiving ACK, Changing state to ESTABLISHED If a packet is lost or dropped between 1 and 6, the client re-sends a packet again But the client waits three seconds to resend, which sacrifices response times
Database connections Web Server Web Server Web Server Web Server Web Server Web Server Web Server Master/Slave Web Server Non-Persistent connections are majority Right now it’s not so much issue – Establishing / terminating network connections is not relatively costly compared to disk access overheads – Waiting 3 seconds for TCP SYN retry is frequent problem Overheads will not be negligible when bottleneck shifts to CPU/Network – Expensive network accesses (SYN-> SYN+ACK -> ACK, initialization SQL statements (SET xxx), etc)
Persistent Connections Web Server Web Server Web Server Web Server Web Server Web Server Web Server Master/Slave Web Server Keeping connections established permanently No connection/disconnection overheads The number of connections equals to “The number of Web Servers” x “Connections per process” Easily reaches 10000 connections max-connections my.cnf has to be increased It’s effective as long as the number of web servers is small… It doesn’t work when the number of web servers becomes large Performance doesn’t drop even though there are thousands of idle connections But RAM space is consumed If mysqld stalls for a short period of time (0.5-1 second), suddenly 1000+ clients execute queries at the same time Very serious problems such as OOM might happen
Use persistent connections properly Queue MySQL Web Worker Servers Servers Servers Program (Q4M) (InnoDB) Uncontrollable number of Limited Number of Non-Persistent Connections Persistent Connections If the number of connections can be controlled, persistent connection helps a lot Before using persistent connections between queue servers and MySQL servers, very often the number of connections was suddenly increased, which caused some errors (1023 error, deadlock error, etc)
Proxy-based connection pooling Web Server Web Server Web Server Web Server Proxy Server Web Server Web Server Web Server Master/Slave Web Server Web servers establish persistent connections to a proxy server, and Proxy server establishes limited number of persistent connections to MySQL Similar to Java EE Proxy server is a lightweight network server It’s not serious even though the number of connections reach 10,000 The number of connections at MySQL is limited, so memory usage is ok, but.. Proxy server has to receive/send massive network requests Response time increases Need High Availability architecture for Proxy
Thread pooling within MySQL Web Server Web Server Web Server Web Server Thread Pool Web Server Web Server Web Server Master/Slave Web Server For stability reasons, the number of threads within mysqld should be limited Adding proxy servers costs more, especially when database servers become faster It is better that mysqld accepts thousands of connections but running limited number of threads internally: Pool of threads MySQL has thread pool plugin interface
Useful commands and tools iostat mpstat dstat oprofile gdb pmp (Poor Man’s Profiler) gcore
iostat Showing detailed I/O statistics per device Very important tool because in most cases RDBMS becomes I/O bound iostat -x Check r/s, w/s, svctm, %util IOPS is much more important than transfer size Always %util = (r/s + w/s) * svctm (hard coded in the iostat source file)# iostat -xm 10avg-cpu:avg-cpu: %user %nice %system %iowait %steal %idle %iowait 21.16 0.00 6.14 29.77 0.00 42.93Device: rqm/s wrqm/s r/s w/s rMB/s wMB/s avgrq-sz avgqu-sz await svctm %util avgrq- avgqu- sdb 2.60 389.01 283.12 47.35 4.86 2.19 43.67 4.89 14.76 3.02 99.83 (283.12+47.35) * 3.02(ms)/1000 = 0.9980 = 100% util
iostat example (DBT-2)# iostat -xm 10avg-cpu:avg-cpu: %user %nice %system %iowait %steal %idle %iowait 21.16 0.00 6.14 29.77 0.00 42.93Device: rqm/s wrqm/s r/s w/s rMB/s wMB/s avgrq-sz avgqu-sz await svctm %util avgrq- avgqu- sdb 2.60 389.01 283.12 47.35 4.86 2.19 43.67 4.89 14.76 3.02 99.83 (283.12+47.35) * 3.02(ms)/1000 = 0.9980 = 100% util# iostat -xm 10avg-cpu:avg-cpu: %user %nice %system %iowait %steal %idle %iowait 40.03 0.00 16.51 16.52 0.00 26.94Device: rrqm/s wrqm/s r/s avgrq- avgqu- w/s rMB/s wMB/s avgrq-sz avgqu-sz await svctm %util sdb 6.39 368.53 543.06 490.41 6.71 3.90 21.02 3.29 3.20 0.90 92.66 3.20 (543.06+490.41) * 0.90(ms)/1000 = 0.9301 = 93% util Sometimes throughput gets higher even though %util reaches 100% Write cache, Command Queuing, etc In both cases %util is almost 100%, but r/s and w/s are far different Do not trust %util too much Check svctm rather than %util If your storage can handle 1000 IOPS, svctm should be less than 1.00 (ms) so you can send alerts if svctm is higher than 1.00 for a couple of minutes
mpstat Per CPU core statistics vmstat displays average statistics It’s very commom that only one of CPU cores consumes 100% CPU resources The rest CPU cores are idle Especially applies to batch jobs If you check only vmstat/top/iostat/sar you will not notice single threaded bottleneck You can also check network bottlenecks (%irq, %soft) from mpstat vmstat counts them as %idle
vmstat and mpstat# vmstat 1…procs -----------memory---------- ---swap-- -----io---- --system-- -----cpu------ -----------memory memory---------- ---swap -----io swap-- io---- --system-- -----cpu system cpu------ r b swpd free buff cache si so bi bo in cs us sy id wa st 0 1 2096472 1645132 18648 19292 0 0 4848 517 0 1223 517 0 0 88 12 0 0 1 2096472 1645132 18648 19292 0 0 4176 0 1287 623 0 0 87 12 0 0 1 2096472 1645132 18648 19292 0 0 4320 0 1202 470 0 0 88 12 0 0 1 2096472 1645132 18648 19292 0 0 3872 0 1289 627 0 0 87 12 0# mpstat -P ALL 1...11:04:37 AM CPU %user %nice %sys %iowait %irq %soft %steal %idle intr/s11:04:38 AM all 0.00 0.00 0.12 12.33 0.00 0.00 0.00 87.55 1201.9811:04:38 AM 0 0.00 0.00 0.00 0.00 0.00 0.00 0.00 100.00 990.1011:04:38 AM 1 0.00 0.00 0.00 0.00 0.00 0.00 0.00 100.00 0.0011:04:38 AM 2 0.00 0.00 0.00 0.00 0.00 0.00 0.00 100.00 0.0011:04:38 AM 3 0.00 0.00 0.00 0.00 0.00 0.00 0.00 100.00 0.0011:04:38 AM 4 0.99 0.00 0.99 98.02 0.00 0.00 0.00 0.00 206.9311:04:38 AM 5 0.00 0.00 0.00 0.00 0.00 0.00 0.00 100.00 0.0011:04:38 AM 6 0.00 0.00 0.00 0.00 0.00 0.00 0.00 100.00 4.9511:04:38 AM 7 0.00 0.00 0.00 0.00 0.00 0.00 0.00 100.00 0.00 vmstat displays average statistics. 12% * 8 (average) = 100% * 1 + 0% * 7
Typical mpstat output Replication Slave %idle on a single CPU core becomes 0-10% – Mostly used for %user, %system, %iowait, %soft Bulk loading/inserting on InnoDB %idle on a single CPU core becomes 0-10% – Mostly used for %user, and %iowait
dstat # dstat ----total-cpu-usage---- -dsk/total- -net/total- ---paging-- ---system-- usr sys idl wai hiq siq| read writ| recv send| in out | int csw 3 1 95 0 0 0|2096k 13M|1437k 2096k| 0 0 |5644 11k 3 2 95 0 0 0|1752k 4096k|1555k 2372k| 0 0 |6015 11k 3 1 95 0 0 0|4640k 12M|1689k 2659k| 0 0 |6092 12k 2 1 96 0 0 0|1616k 5352k|1590k 2493k| 0 0 |5593 10k 3 1 96 0 0 0|1968k 6960k|1558k 2400k| 0 0 |5663 10k # dstat -N bond0 ----total-cpu-usage---- -dsk/total- -net/bond0- ---paging-- ---system-- net/bond0- usr sys idl wai hiq siq| read writ| recv send| in out | int csw send 2 1 95 1 0 0|2416k 5112k| 782k 1196k 1196k| 0 0 |5815 11k 3 1 95 0 0 0|3376k 15M| 792k 1213k 1213k| 0 0 |5632 11k 3 2 95 0 0 0|2120k 3264k| 793k 1229k 1229k| 0 0 |5707 11k 2 1 96 0 0 0|1920k 3560k| 788k 1193k 1193k| 0 0 |5856 11k 2 1 96 0 0 0|1872k 13M| 770k 1112k 1112k| 0 0 |5463 10k yum install dstat Similar UI as vmstat, but it also shows network statistics Disk and Net total is incorrect, if you use RAID/bonding (double counted) Filter by disk/interface name that you want to trace “mtstat” additionally supports mysql status outputs https://launchpad.net/mtstat
Oprofile Profiling CPU usage from running processes You can easily identify which functions consume CPU resources Supporting both user space and system space profiling Mainly used by database-internal developers If specific functions consume most of recourses, applications might be re-designed to skip calling them Not useful to check low-CPU activities I/O bound, mutex waits, etc How to use opcontrol --start (--no-vmlinux) benchmarking opcontrol --dump opcontrol --shutdown opreport -l /usr/local/bin/mysqld
Oprofile example /usr/local/bin/mysqld# opreport –l /usr/local/bin/mysqldsamples % symbol name83003 String::copy(char int, charset_info_st*, 8.8858 String::copy(char const*, unsigned int, charset_info_st*,charset_info_st*, int*)charset_info_st*, unsigned int*)79125 8.4706 MYSQLparse(void*) MYSQLparse(void*)68253 7.3067 my_wc_mb_latin155410 5.9318 my_pthread_fastmutex_lock34677 3.7123 my_utf8_uni18359 MYSQLlex(void*, 1.9654 MYSQLlex(void*, void*)12044 1.2894 _ZL15get_hash_symbolPKcjb11425 _ZL20make_join_statisticsP4JOINP10TABLE_LISTP4ItemP16st_dynamic_array 1.2231 _ZL20make_join_statisticsP4JOINP10TABLE_LISTP4ItemP16st_dynamic_array You can see quite a lot of CPU resources were spent for character conversions (latin1 <-> utf8) Disabling character code conversions on application side will improve performancesamples % symbol name83107 MYSQLparse(void*) 10.6202 MYSQLparse(void*)68680 8.7765 my_pthread_fastmutex_lock20469 MYSQLlex(void*, 2.6157 MYSQLlex(void*, void*)13083 1.6719 _ZL15get_hash_symbolPKcjb12148 JOIN::optimize() 1.5524 JOIN::optimize()11529 _ZL20make_join_statisticsP4JOINP10TABLE_LISTP4ItemP16st_dynamic_array 1.4733 _ZL20make_join_statisticsP4JOINP10TABLE_LISTP4ItemP16st_dynamic_array
Checking stalls Database servers sometimes suffer from performance stalls Not responding anything for 1-2 seconds In many cases it’s database internal issue Holding global mutex and doing expensive i/o Buffer pool, redo log preflush, rollback segments (InnoDB mutex) pthread_create()/clone() (LOCK_thread_count mutex) Opening/closing tables (LOCK_open mutex) Collecting statistics Per-minute statistics (i.e. vmstat 60) is not helpful. Collect per-second statistics
Checking stalls by gdb or pmp Debugging tool gdb has a functionality to take thread stack dumps from a running process Useful to identify where and why mysqld hangs up, slows down, etc But you have to read MySQL source code During taking stack dumps, all threads are stopped Debugging symbol is required on the target program
gdb case study mysql> mysql> SELECT query_time, start_time, sql_text query_time, start_time, -> FROM mysql.slow_log WHERE start_time -> BETWEEN 2010-02-05 23:00:00 AND 2010-02-05 01:00:00 2010-02- 2010-02- -> ORDER BY query_time DESC LIMIT 10; ------------+---------------------+----------+ +------------+---------------------+----------+ Suddenly all queries | query_time | start_time | sql_text | were not responding for ------------+---------------------+----------+ +------------+---------------------+----------+ 1-10 seconds | 00:00:11 | 2010-02-05 23:09:55 | begin 2010-02- | | 00:00:09 | 2010-02-05 23:09:55 | Prepare | 2010-02- Checking slow query log | 00:00:08 | 2010-02-05 23:09:55 | Prepare | 2010-02- All queries are simple 2010-02- | 00:00:08 | 2010-02-05 23:09:55 | Init DB | | 00:00:08 | 2010-02-05 23:09:55 | Init DB | 2010-02- enough, it’s strange to 2010-02- | 00:00:07 | 2010-02-05 23:09:55 | Prepare | take 10 seconds 2010-02- | 00:00:07 | 2010-02-05 23:09:55 | Init DB | CPU util (%us, %sy) 2010-02- | 00:00:07 | 2010-02-05 23:09:55 | Init DB | | 00:00:07 | 2010-02-05 23:09:55 | Init DB | 2010-02- were almost zero | 00:00:06 | 2010-02-05 23:09:55 | Prepare | 2010-02- SHOW GLOBAL ------------+---------------------+----------+ +------------+---------------------+----------+ STATUS, SHOW FULL 10 rows in set (0.02 sec) PROCESSLIST were not helpful
Taking thread dumps with gdb gdbtrace() { … PID=`cat /var/lib/mysql/mysql.pid` STACKDUMP=/tmp/stackdump.$$ Attaching running echo thread apply all bt > thread bt mysqld, then taking a $STACKDUMP thread dump echo detach >> $STACKDUMP detach detach Taking dumps every 3 echo quit >> $STACKDUMP quit quit seconds gdb --batch --pid=$PID -x $STACKDUMP --batch --pid pid=$PID Attaching & Dumping } with gdb is expensive so while loop invoke only when do exceptional scenario (i.e. CONN=`netstat -an | grep 3306 | grep conn > threshold) ESTABLISHED | wc | awk {print $1}` happens if [ $CONN -gt 100 ]; then gdbtrace() Check if the same LWPs done are waiting at the same sleep 3 done place
Stack Trace.....Thread 73 (Thread 0x46c1d950 (LWP 28494)):#0 0x00007ffda5474384 in __lll_lock_wait () from /lib/libpthread.so.0#1 0x00007ffda546fc5c in _L_lock_1054 () from /lib/libpthread.so.0#2 0x00007ffda546fb30 in pthread_mutex_lock () from /lib/libpthread.so.0#3 0x0000000000a0f67d in my_pthread_fastmutex_lock (mp=0xf46d30) atthr_mutex.c:487#4 0x000000000060cbe4 in dispatch_command (command=16018736, thd=0x80, packet=0x65 <Address 0x65 out of bounds>, packet_length=4294967295) at sql_parse.cc:969#5 0x000000000060cb56 in do_command (thd=0xf46d30) at sql_parse.cc:854#6 0x0000000000607f0c in handle_one_connection (arg=0xf46d30) atsql_connect.cc:1127#7 0x00007ffda546dfc7 in start_thread () from /lib/libpthread.so.0#8 0x00007ffda46305ad in clone () from /lib/libc.so.6#9 0x0000000000000000 in ?? () Many threads were waiting at pthread_mutex_lock(), called from sql_parse.cc:969
Reading sql_parse.cc:969 *thd thd, 953 bool dispatch_command(enum enum_server_command command, THD *thd, 954 char* packet, uint packet_length) packet_length) 955 { 956 NET *net= &thd->net; *net &thd net= thd- 957 bool error= 0; 958 DBUG_ENTER("dispatch_command"); DBUG_ENTER("dispatch_command"); 959 DBUG_PRINT("info",("packet: packet_length, DBUG_PRINT("info",("packet: %*.s; command: %d", packet_length, packet, command)); 960 961 thd->command=command; thd->command=command; command 962 /* 963 Commands which always take a long time are logged into 964 the slow log only if opt_log_slow_admin_statements is set. 965 */ 966 thd- enable_slow_log= thd->enable_slow_log= TRUE; 967 thd- lex- sql_command= thd->lex->sql_command= SQLCOM_END; /* to avoid confusing VIEW detectors */ 968 thd- set_time(); thd->set_time(); 969 VOID(pthread_mutex_lock(&LOCK_thread_count)); VOID(pthread_mutex_lock(&LOCK_thread_count));
Who locked LOCK_thread_count for seconds?Thread 1 (Thread 0x7ffda58936e0 (LWP 15380)): LWP 15380#0 0x00007ffda4630571 in clone () from /lib/libc.so.6#1 0x00007ffda546d396 in do_clone () from /lib/libpthread.so.0#2 0x00007ffda546db48 in pthread_create@@GLIBC_2.2.5 () from/lib/libpthread.so.0#3 0x0000000000600a66 in create_thread_to_handle_connection (thd=0x3d0f00) at mysqld.cc:4811#4 0x00000000005ff65a in handle_connections_sockets (arg=0x3d0f00) atmysqld.cc:5134#5 0x00000000005fe6fd in main (argc=4001536, argv=0x4578c260) atmysqld.cc:4471#0 0x00007ffda4630571 in clone () from /lib/libc.so.6 gdb stack dumps were taken every 3 seconds In all cases, Thread 1 (LWP 15380) was stopped at the same point clone() (called by pthread_create()) seemed to take a long time
Reading mysqld.cc:4811 thd) 4795 void create_thread_to_handle_connection(THD *thd) 4796 { (cached_thread_count wake_thread) 4797 if (cached_thread_count > wake_thread) 4798 { 4799 /* Get thread from cache */ 4800 thread_cache.append(thd); thread_cache.append(thd); 4801 wake_thread++; wake_thread++; 4802 pthread_cond_signal(&COND_thread_cache); pthread_cond_signal(&COND_thread_cache); 4803 } 4804 else 4805 { 4811 ((error=pthread_create(&thd- real_id,&connection_attrib, pthread_create(&thd if ((error=pthread_create(&thd->real_id,&connection_attrib, 4812 handle_one_connection, handle_one_connection, 4813 thd))) (void*) thd))) 4839 } 4840 (void) pthread_mutex_unlock(&LOCK_thread_count); pthread_mutex_unlock(&LOCK_thread_count); pthread_create is called under critical section (LOCK_thread_count is released after that) If cached_thread_count > wake_thread, pthread_create is not called Increasing thread_cache_size will fix the problem!
Poor Man’s Profiler A simple shell script (wrapper script) to summarize thread call stacks Using gdb thread dump functionality Useful for identifying why MySQL stalls http://poormansprofiler.org/ #!/bin/bash nsamples=1 sleeptime=0 pid=$(pidof mysqld) for x in $(seq 1 $nsamples) do gdb -ex "set pagination 0" -ex "thread apply all bt" -batch -p $pid sleep $sleeptime done | ¥ awk BEGIN { s = ""; } /Thread/ { print s; s = ""; } /^¥#/ { if (s != "" ) { s = s "," $4} else { s = $4 } } END { print s } | ¥ sort | uniq -c | sort -r -n -k 1,1
pmp output example291 pthread_cond_wait@@GLIBC_2.3.2,one_thread_per_connection_end,handle_one_connection 57 read,my_real_read,my_net_read,do_command,handle_one_connection,start_thread 26pthread_cond_wait@@GLIBC_2.3.2,os_event_wait_low,os_aio_simulated_handle,fil_aio_wait,io_handler_thread,start_thread 3 pthread_cond_wait@@GLIBC_2.3.2,os_event_wait_low,srv_purge_worker_thread 1 select,os_thread_sleep,srv_purge_thread 1 select,os_thread_sleep,srv_master_thread 1 select,os_thread_sleep,srv_lock_timeout_and_monitor_thread 1 select,os_thread_sleep,srv_error_monitor_thread 1 select,handle_connections_sockets,main,select 1 read,vio_read_buff,my_real_read,my_net_read,cli_safe_read,handle_slave_io 1pthread_cond_wait@@GLIBC_2.3.2,os_event_wait_low,sync_array_wait_event,rw_lock_s_lock_spin,buf_page_get_gen,btr_cur_search_to_nth_level,row_search_for_mysql,ha_innodb::index_read,handler::index_read_idx_map,join_read_const,join_read_const_table,make_join_statistics,JOIN::optimize,mysql_select,handle_select,execute_sqlcom_select,mysql_execute_command,mysql_parse,dispatch_command,do_command,handle_one_connection
Disadvantages of gdb/pmp All threads are stopped while taking thread dumps The more the number of threads, the more time it takes It might take 5 seconds if 1000 threads are running 5 seconds stall is not acceptable in many cases It is not recommended taking thread dumps on master when lots of threads are running Be careful if you set high thread_cache_size
gcore Dumping core file Dumping core file takes time, and blocks all threads during dump Might take 1 hour when VIRT is 50GB+ Not useful on running systems, but can be used on a non- serving slave If you have a program with DWARF symbols, target (running) process does not need symbols You can run stripped mysqld on production, as long as you have debug-built mysqld
Case studyProblem:- mysqld on one of slaves used much more CPU resources than other slaves,and frequently replication delay happened- the mysqld used 50GB+ VIRT on 24G box. (other slaves used 20G VIRT) top - 20:39:14 up 360 days, 17:56, 1 user, load average: 1.26, 1.29, 1.32 Tasks: 125 total, 2 running, 123 sleeping, 0 stopped, 0 zombie Cpu(s): 10.8% us, 0.8% sy, 0.0% ni, 87.2% id, 0.7% wa, 0.0% hi, 0.5% si Mem: 24680588k total, 24609132k used, 71456k free, 99260k buffers Swap: 4192956k total, 160744k used, 4032212k free, 4026256k cached PID USER PR NI VIRT RES SHR S %CPU %MEM TIME+ COMMAND 11706 mysql 16 0 50.5g 19g 5168 R 90.9 81.4 247117:59 mysqld 19253 root 16 0 6172 1164 852 R 0.3 0.0 0:00.03 top
From mysql status Slave_open_temp_tables was extremely high. | Slave_open_temp_tables | 15927 | But we don’t use tmp tables right now. Previously developers created tmp tables, but in theory SQL thread should have closed them. SQL thread internally checks all its tmp tables whether it can be closed or not, per each SQL statement. Parsing 15927 tmp tables per statement was very expensive and it should be the reason that caused replication delay. I wanted to know what tables were opened, but MySQL command line tools do not provide ways to identify them. Time to analyze core file !
Core file analysis example # Install debuginfo mysqld if not installed (don’t need to run debugged mysqld) # ulimit -c unlimited # gdb -p $(pidof mysqld) (gdb) gcore core (Takes long time..) (gdb) detach # gdb /usr/lib/debug/usr/sbin/mysqld.debug /data/mysql/core.11706 (gdb) p active_mi->rli->save_temporary_tables->s->table_name $1 = {str = 0x31da400c3e "claimed_guids", length = 13} (gdb) p $a->file->stats $16 = {data_file_length = 1044496, max_data_file_length = 5592400, index_file_length = 1044496, max_index_file_length = 0, delete_length = 0, auto_increment_value = 0, records = 1, deleted = 0, mean_rec_length = 8, create_time = 0, check_time = 0, update_time = 0, block_size = 0} Tmp tables are MEMORY engine so 1044496 + 1044496 bytes (2MB) were used per table. Table name was “claimed_guids”
Dumping slave’s tmp tables info define print_all_tmp_tables set $a= active_mi->rli->save_temporary_tables set $b= slave_open_temp_tables while ($i < $b) p $a->alias p $a->file->stats set $a= $a->next set $i=$i+1 end end set pagination 0 set $i=0 print_all_tmp_tables detach quit# gdb /usr/lib/debug/usr/sbin/mysqld.debug /data/mysql/core.11706 -x above_script List length was 15927 (same as slave_open_tmp_tables), all tables were the same name, used same size (2MB) 15927 * 2MB = 31GB This should be the reason why VIRT was 30+ GB larger than other slaves. Rebooting the slave should be fine because they are tmp tables.
Capturing MySQL packets Capturing MySQL packets is useful to trace slow queries / transactions / connection requests. Modifying applications MySQL settings is not needed – Starting from 5.1, you can dynamically enable/disable slow/general query logs libpcap library to capture network packets Most of network capturing tools including tcpdump rely on libpcap Packet loss often happens
tcpdump + mk-query-digest1) [root#] tcpdump -i bond0 port 3306 -s 65535 -x -n -q -tttt -c 10000 > tcpdump.out2) $ mk-query-digest --type=tcpdump --report-format=query_report tcpdump.out# Query 1: 951.34 QPS, 1.56x concurrency, ID 0x6D23B63A5DA67F5D at byte# Attribute pct total min max avg 95% stddev median# ========= ====== ======= ======= ======= ======= ======= =======# Count 22 1837# Exec time 40 3s 0 207ms 2ms 8ms 6ms 236usSELECT * FROM user WHERE user_id = 100;# Query 2: 1.06k QPS, 0.69x concurrency, ID 0x2EB286C018F2DE27 at byte mk-query-digest has a functionality to parse tcpdump outputs It can trace how long each query takes, and it can summarize and sort by total execution time No need to modify my.cnf settings Queries that take less than 1 second can be captured Very helpful in MySQL 5.0 or eariler
MySlowTranCapture[mysql-server]# myslowtrancapture -i eth0Monitoring eth0 interface..Listening port 3306.. From 192.168.0.2:24442Transactions that take more than 4000 2011/01/07 09:12:20.969288 ->milliseconds.. begin 2011/01/07 09:12:20.969483 <-From 192.168.0.1:24441 GOT_OK2011/01/07 09:12:17.258307 -> 2011/01/07 09:12:20.977699 ->begin update diary set diary_date=now() where2011/01/07 09:12:17.258354 <- diary_id=100GOT_OK 2011/01/07 09:13:11.300935 <-2011/01/07 09:12:17.264797 -> GOT_ERR:Lock wait timeout exceeded; tryselect * from diary where diary_id=100 for restarting transactionupdate 2011/01/07 09:13:13.136967 ->2011/01/07 09:12:17.265087 <- rollbackGOT_RES2011/01/07 09:13:01.232620 ->update diary set diary_date=now() wherediary_id=1002011/01/07 09:13:01.232960 <- https://github.com/yoshinorim/MySlowTranCaptureGOT_OK A tool to print “transactions” that take more than N milliseconds2011/01/07 09:13:17.360993 -> Held locks are not released until COMMIT/ROLLBACK is executedcommit
Summary
SSD SATA SSD Much more cost effective than spending on RAM only H/W RAID Controller + SATA SSD won’t perform great Capacitor + SSD should be great. Check Intel 320 PCI-Express SSD Check FusionIO and tachIOn Expensive, but MLC might be affordable Master handles traffics greatly, but slave servers can not catch up due to single threaded replication channel Consider running multiple instances on a slave server
RAM Allocate swap space (approx half of RAM size) Set vm.swappiness = 0 and use O_DIRECT Restart linux if kernel panic happens – kernel.panic_on_oops = 1 – kernel.panic = 1 Use linux kernel 2.6.28+, or frequently unmap logs from filesystem cache Do not allocate too much memory per session Do not use persistent connections if you can’t control maximum number of MySQL internal threads

posted on 2011-10-14 14:44 zhizhesky 阅读(818) 评论(0) 编辑收藏举报

会员力量，点亮园子希望

刷新页面返回顶部

半夜ATM机前看书的那位

导航

公告

Linux and H/W optimizations for MySQL