在 Linux 操作系统上运行内存需求量较大的应用程序时,由于其采用的默认页面大小为 4KB,因而将会产生较多 TLB Miss 和缺页中断,从而大大影响应用程序的性能。当操作系统以 2MB 甚至更大作为分页的单位时,将会大大减少 TLB Miss 和缺页中断的数量,显著提高应用程序的性能。这也正是 Linux 内核引入大页面支持的直接原因。好处是很明显的,假设应用程序需要 2MB 的内存,如果操作系统以 4KB 作为分页的单位,则需要 512 个页面,进而在 TLB 中需要 512 个表项,同时也需要 512 个页表项,操作系统需要经历至少 512 次 TLB Miss 和 512 次缺页中断才能将 2MB 应用程序空间全部映射到物理内存;然而,当操作系统采用 2MB 作为分页的基本单位时,只需要一次 TLB Miss 和一次缺页中断,就可以为 2MB 的应用程序空间建立虚实映射,并在运行过程中无需再经历 TLB Miss 和缺页中断(假设未发生 TLB 项替换和 Swap)。
为了能以最小的代价实现大页面支持,Linux 操作系统采用了基于 hugetlbfs 特殊文件系统 2M 字节大页面支持。这种采用特殊文件系统形式支持大页面的方式,使得应用程序可以根据需要灵活地选择虚存页面大小,而不会被强制使用 2MB 大页面。
在mysql中,innodb可以使用large pages来分配buffer pool跟additional memory
pool.
下面来看下mysql使用large pages具体的实现方法
1:查看系统是否支持huge-pages
[root@localhost ~]# cat /proc/meminfo | grep -i huge AnonHugePages: 47708160 kB HugePages_Total: 0 HugePages_Free: 0 HugePages_Rsvd: 0 HugePages_Surp: 0 Hugepagesize: 2048 kB [root@localhost ~]#
Hugepagesize有值,说明系统支持large pages(如不支持,系统需要重新编译来支持)
2.修改预分配的hugepages大小
[root@localhost ~]# sysctl -w vm.nr_hugepages=512 vm.nr_hugepages = 512 [root@localhost ~]#
[root@localhost ~]# cat /proc/meminfo | grep -i huge AnonHugePages: 79872 kB HugePages_Total: 512 HugePages_Free: 505 HugePages_Rsvd: 229 HugePages_Surp: 0 Hugepagesize: 2048 kB [root@localhost ~]#
3.查看mysql用户的组信息
[root@localhost ~]# id mysql uid=500(mysql) gid=501(mysql) groups=501(mysql) [root@localhost ~]#
4.配置使用大页内存的用户组
[root@localhost ~]# sysctl -w vm.hugetlb_shm_group=501 vm.hugetlb_shm_group = 501 [root@localhost ~]#
5.修改ulimit
vim /etc/security/limits.conf
@mysql soft memlock unlimited
@mysql hard memlock unlimited
6.修改/etc/sysconfig.conf
# Increase the amount of shmem allowed per segment # This depends upon your memory, remember your kernel.shmmax = 68719476736 # Increase total amount of shared memory. kernel.shmall = 4294967296
shmmax 是最大的共享内存段的大小,单位是字节,默认32M,肯定是不够的,这个应该比innodb_buffer_pool要大。shmall是共享内存的总大小,单位是页,默认2097152(8G)。可以使用sysctl -w或者在/etc/sysctl.conf中设置。
sysctl -p使其设置生效。
7.修改my.cnf
[mysqld]
large-pages
8.重启mysql
[root@localhost mysql]# /etc/init.d/mysqld restart Shutting down MySQL. [ OK ] Starting MySQL.... [ OK ] [root@localhost mysql]# cat localhost.localdomain.err 140320 15:18:12 [Note] /usr/local/mysql/bin/mysqld: Normal shutdown 140320 15:18:12 [Note] Event Scheduler: Purging the queue. 0 events 140320 15:18:12 InnoDB: Starting shutdown... 140320 15:18:12 InnoDB: Shutdown completed; log sequence number 3629001490 140320 15:18:12 [Note] /usr/local/mysql/bin/mysqld: Shutdown complete 140320 15:18:12 mysqld_safe mysqld from pid file /data/mysql/localhost.localdomain.pid ended 140320 15:18:13 mysqld_safe Starting mysqld daemon with databases from /data/mysql 140320 15:18:13 InnoDB: The InnoDB memory heap is disabled 140320 15:18:13 InnoDB: Mutexes and rw_locks use GCC atomic builtins 140320 15:18:13 InnoDB: Compressed tables use zlib 1.2.3 140320 15:18:13 InnoDB: Initializing buffer pool, size = 256.0M 140320 15:18:13 InnoDB: Completed initialization of buffer pool 140320 15:18:13 InnoDB: highest supported file format is Barracuda. 140320 15:18:15 InnoDB: Waiting for the background threads to start 140320 15:18:16 InnoDB: 1.1.8 started; log sequence number 3629001490 140320 15:18:16 [Note] Server hostname (bind-address): ‘0.0.0.0‘; port: 3306 140320 15:18:16 [Note] - ‘0.0.0.0‘ resolves to ‘0.0.0.0‘; 140320 15:18:16 [Note] Server socket created on IP: ‘0.0.0.0‘. 140320 15:18:16 [Note] Event Scheduler: Loaded 0 events 140320 15:18:16 [Note] /usr/local/mysql/bin/mysqld: ready for connections. Version: ‘5.5.25a-log‘ socket: ‘/tmp/mysqld.sock‘ port: 3306 Source distribution
可以看见没有报错,假如我们调小nr_hugepages为200看看
[root@localhost mysql]# sysctl -w vm.nr_hugepages=200 vm.nr_hugepages = 200 [root@localhost mysql]# /etc/init.d/mysqld restart Shutting down MySQL. [ OK ] Starting MySQL... [ OK ] [root@localhost mysql]# cat localhost.localdomain.err 140320 15:20:08 [Note] /usr/local/mysql/bin/mysqld: Normal shutdown 140320 15:20:08 [Note] Event Scheduler: Purging the queue. 0 events 140320 15:20:08 InnoDB: Starting shutdown... 140320 15:20:08 InnoDB: Shutdown completed; log sequence number 3629001490 140320 15:20:08 [Note] /usr/local/mysql/bin/mysqld: Shutdown complete 140320 15:20:08 mysqld_safe mysqld from pid file /data/mysql/localhost.localdomain.pid ended 140320 15:20:09 mysqld_safe Starting mysqld daemon with databases from /data/mysql 140320 15:20:09 InnoDB: The InnoDB memory heap is disabled 140320 15:20:09 InnoDB: Mutexes and rw_locks use GCC atomic builtins 140320 15:20:09 InnoDB: Compressed tables use zlib 1.2.3 140320 15:20:09 InnoDB: Initializing buffer pool, size = 256.0M InnoDB: HugeTLB: Warning: Failed to allocate 274726912 bytes. errno 12 InnoDB HugeTLB: Warning: Using conventional memory pool 140320 15:20:09 InnoDB: Completed initialization of buffer pool 140320 15:20:09 InnoDB: highest supported file format is Barracuda. 140320 15:20:11 InnoDB: Waiting for the background threads to start 140320 15:20:12 InnoDB: 1.1.8 started; log sequence number 3629001490 140320 15:20:12 [Note] Server hostname (bind-address): ‘0.0.0.0‘; port: 3306 140320 15:20:12 [Note] - ‘0.0.0.0‘ resolves to ‘0.0.0.0‘; 140320 15:20:12 [Note] Server socket created on IP: ‘0.0.0.0‘. 140320 15:20:12 [Note] Event Scheduler: Loaded 0 events 140320 15:20:12 [Note] /usr/local/mysql/bin/mysqld: ready for connections. Version: ‘5.5.25a-log‘ socket: ‘/tmp/mysqld.sock‘ port: 3306 Source distribution [root@localhost mysql]#
上面出现了如下警告:
InnoDB: HugeTLB: Warning: Failed to allocate 274726912 bytes. errno 12 InnoDB HugeTLB: Warning: Using conventional memory pool
因为hugepage分配内存的时候,是一次性、且独占的, 一次性指在mysqld起来的时候,所有buffer
pool需要的内存总和一次性的被分配,而且这些分配的内存不能被其他进程占用。所以就是一次性且独占。而当buffer
pool所需要的内存被一次性的分配了之后,那么必然就不会使用swap了。
using conventional memory
pool是什么意思呢?因为你的hugepage内存不够你的buffer pool大小,所以就用常规内存了。
开启大页内存的好处:
1.减少内存置换
2.减少TLB miss次数
3.减少swap
在启动mysql的时候,最容易报的错误是:
InnoDB: HugeTLB: Warning: Failed to allocate 274726912 bytes. errno
12
InnoDB HugeTLB: Warning: Using conventional memory
pool
这是由于上面配置的两点所致。
1)nr_hugepages
的值*2M应大于(innodb_buffer_pool_size+innodb_additional_mem_pool_size)因为以上谈到
innodb可以使用large pages来分配buffer pool跟additional memory
pool.
2)memlock
的设置,在启动mysql时,一定要先查看用ulimit -a 来查看max locked memory
设置是否合理,可以尝试用以上两种方法来设置该值。还有一点,/etc/security/limits.conf配置文件的修改.