MySQL 单表插入 10w+ TPS达成 -k6k4.com

装B留念:

MySQL服务端机器配置和操作系统信息,没有使用FlashCache哦:

# Aspersa System Summary Report ##############################Uptime | 84 days, 19:00, 3 users, load average: 0.00, 0.21, 0.16

Platform | Linux

Release | Red Hat Enterprise Linux Server release 5.4 (Tikanga)

Kernel | 2.6.18-164.el5

Architecture | CPU = 64-bit, OS = 64-bit

Threading | NPTL 2.5

Compiler | GNU CC version 4.1.2 20080704 (Red Hat 4.1.2-44).

SELinux | Disabled

Virtualized | No virtualization detected

# Processor ##################################################

Processors | physical = 2, cores = 8, virtual = 16, hyperthreading = yes

Speeds | 16x2261.058

Models | 16xIntel(R) Xeon(R) CPU E5520 @ 2.27GHz

Caches | 16x8192 KB

# Memory #####################################################

Total | 23.53G

Free | 2.16G

Used | physical = 21.38G, swap = 240.00k, virtual = 21.38G

Buffers | 1.03G

Caches | 13.60G

Dirty | 156 kB

UsedRSS | 6.1G

Swappiness | vm.swappiness = 60

DirtyPolicy | vm.dirty_ratio = 40, vm.dirty_background_ratio = 10

# Mounted Filesystems ########################################

Filesystem Size Used Type Opts Mountpoint

/dev/sda10 766G 11% ext3 rw /uxx

/dev/sda1 122M 16% ext3 rw /boot

/dev/sda2 15G 67% ext3 rw /

/dev/sda3 15G 76% ext3 rw /usr

/dev/sda5 8.6G 2% ext3 rw /tmp

tmpfs 12G 0% tmpfs rw /dev/shm

# Disk Schedulers And Queue Size #############################

sda | [cfq] 128

# Disk Partioning ############################################

Device Type Start End Size

============ ==== ========== ========== ==================

# Kernel Inode State #########################################

dentry-state | 297447 276749 45 0 0 0

file-nr | 3570 0 2390094

inode-nr | 220730 32

# LVM Volumes ################################################

WARNING: Running as a non-root user. Functionality may be unavailable.

# RAID Controller ############################################

Controller | LSI Logic MegaRAID SAS

Model | , interface, ports

Cache | Memory, BBU

BBU | % Charged, Temperature C, isSOHGood=

VirtualDev Size RAID Level Disks SpnDpth Stripe Status Cache

========== ========= ========== ===== ======= ====== ======= =========

PhysiclDev Type State Errors Vendor Model Size

========== ==== ======= ====== ======= ============ ===========

# Network Config #############################################

Controller | Broadcom Corporation NetXtreme II BCM5709 Gigabit Ethernet (rev 20)

Controller | Broadcom Corporation NetXtreme II BCM5709 Gigabit Ethernet (rev 20)

Controller | Broadcom Corporation NetXtreme II BCM5709 Gigabit Ethernet (rev 20)

Controller | Broadcom Corporation NetXtreme II BCM5709 Gigabit Ethernet (rev 20)

Controller | Broadcom Corporation NetXtreme II BCM5709 Gigabit Ethernet (rev 20)

Controller | Broadcom Corporation NetXtreme II BCM5709 Gigabit Ethernet (rev 20)

Controller | Broadcom Corporation NetXtreme II BCM5709 Gigabit Ethernet (rev 20)

Controller | Broadcom Corporation NetXtreme II BCM5709 Gigabit Ethernet (rev 20)

FIN Timeout | net.ipv4.tcp_fin_timeout = 60

Port Range | net.ipv4.ip_local_port_range = 32768 61000

# Interface Statistics #######################################

interface rx_bytes rx_packets rx_errors tx_bytes tx_packets tx_errors

========= ========= ========== ========== ========== ========== ==========

lo 600000000 500000 0 600000000 500000 0

eth0 0 0 0 0 0 0

eth1 0 0 0 0 0 0

eth2 0 0 0 0 0 0

eth3 0 0 0 0 0 0

eth4 1000000000 600000000 0 2000000000 450000000 0

eth5 0 0 0 0 0 0

eth6 1250000000 15000000 0 0 0 0

eth7 0 0 0 0 0 0

sit0 0 0 0 0 0 0

bond0 2500000000 600000000 0 2000000000 450000000 0

# The End ####################################################

MySQL 使用Percona 5.5.18
my.cnf的配置如下

default-storage-engine = INNODBinnodb_flush_method = O_DIRECT

innodb_file_per_table = 1

innodb_flush_log_at_trx_commit = 1

innodb_lock_wait_timeout = 100

innodb_additional_mem_pool_size = 20M

innodb_buffer_pool_size = 5G

innodb_log_buffer_size= 800M

innodb_log_file_size = 200M

innodb_log_files_in_group = 4

innodb_file_io_threads = 4

innodb_thread_concurrency = 32

innodb_max_dirty_pages_pct = 90

innodb_data_file_path = ibdata1:1G:autoextend

innodb_sync_spin_loops = 0

innodb_spin_wait_delay = 0

tdh_socket_thread_num = 8

tdh_socket_slow_read_thread_num = 64

tdh_socket_io_thread_num = 4

tdh_socket_write_thread_num = 16

tdh_socket_optimize_on = 1

tdh_socket_cache_table_num_for_thd = 40

插入的表结构

CREATE TABLE test (id INT(20) UNSIGNED NOT NULL AUTO_INCREMENT,

k INT(20) DEFAULT NULL,

i INT(20) NOT NULL,

c CHAR(120) DEFAULT NULL,

kc INT(20) DEFAULT ‘1’,

PRIMARY KEY (id)

) ENGINE=InnoDB DEFAULT CHARSET=gbk;

压测脚本:

package benchmark.insertimport benchmark.StressTest

import com.taobao.tdhs.client.TDHSClient

import com.taobao.tdhs.client.TDHSClientImpl

import com.taobao.tdhs.client.response.TDHSResponse

import com.taobao.tdhs.client.response.TDHSResponseEnum

import java.util.concurrent.atomic.AtomicLong

/* @author <a href=”mailto:wentong@taobao.com”>文通</a>

@since 12-2-9 下午3:31 

/

int step = 100000000

int index = Integer.valueOf(args[0])

AtomicLong id = new AtomicLong((index - 1) step + 1)

TDHSClient client = new TDHSClientImpl(new InetSocketAddress(“t-wentong”, 9999), 2);

def s = new StressTest(count: step)

s.add({

def _id = id.incrementAndGet()

String table = “test”

TDHSResponse response = client.createStatement(index).insert().use(“benchmark_insert”).from(table)

.value(“id”, _id.toString())

.value(“k”, “10000”)

.value(“i”, _id.toString())

.value(“c”, _id.toString() + “_abcdefghijklmnopqrstuvwxyz”).insert()

if (response != null && response.getStatus() != TDHSResponseEnum.ClientStatus.OK) {

System.out.println(response);

}

}, 100)

s.run()

client.shutdown()

使用TDH_SOCKET进行插入能到这么高TPS的关键:

TDH_SOCKET提供group commit: 减少redo log的fsync次数,使IOPS不成为瓶颈

TDH_SOCKET能提供并发写. 自增id的生成策略需要分段自增:

如果采用全自增id生成策略(即默认innodb提供的自增id生成策略)的话,会在高并发插入的时候遇到一个block的写锁. 因为是顺序自增,所以并发插入的记录都会集中写入到同一个page上,而一个线程写page会对这个page做一次rw_lock_x_lock_func(&(block->lock), 0, file, line); 那么这个写锁会成为瓶颈,使TPS无法上去所以只要改变生成id的策略:

分段自增比如一个写线程一下子获取1-10000的id范围下一个写线程获取10001-20000的id范围.. 每个写线程在各自的范围里面自增的生成id,那么即能保证一定的顺序写,又能使写page不会因为并发写锁而性能下降!
好,当上了10wTPS之后你会发现CPU占用:

个人资料

0_0
等级：7
文章：111篇
访问：5.0w
排名： 6

推荐圈子

阿里中间件技术交流圈