记录redis "Connection timed out"处理
最近程序报错：Failed to connect to redis: Connection timed out
发现THP的有关信息，决定做测试及修改
1.page 与 Huge Pages
page：
一般而言，内存管理的最小块级单位叫做page，一个page是4096bytes，1M的内存会有256个page，1GB的话就会有256,000个page。
CPU通过内置的内存管理单元维护着page表记录。
Huge pages：
Huge Pages就是大小为2M到1GB的内存page，主要用于管理数千兆的内存，比如1GB的page对于1TB的内存来说是相对比较合适的。
HugePage广泛启用开始于Kernal 2.6,一些版本下2.4内核也可以是用。在操作系统Linux环境中，内存是以页Page的方式进行分配，默认大小为4K。如果需要比较大的内存空间，则需要进行频繁的页分配和管理寻址动作。
HugePage是传统4K Page的替代方案。顾名思义，是用HugePage可以让我们有更大的内存分页大小。无论是HugePage还是传统的正常Page，这个过程都涉及到OS内存寻址过程
当一个进程访问内存的时候，并不是直接进行内存位置访问，是需要通过Page Table进行转移变换。在使用HugePage的情况下，PageTable具有了额外的属性，就是判断该页记录是HugePage还是Regular Page。
2.THP
THP（Transparent Huge Pages）是一个使管理Huge Pages自动化的抽象层。
Linux本身的页大小是固定的4KB，在2.6.38内核新增了THP，透明地支持huge page（2MB）的使用，并且默认开启。
THP优点：
(1) 减少page fault。一次page fault可以加载更大的内存块。
(2) 更小的页表。相同的内存大小，需要更少的页。
(3) 由于页表更小，虚拟地址到物理地址的翻译也更快。
THP缺点：
(1)降低分配内存效率。需要大块、连续内存块，内核线程会比较激进的进行compaction，解决内存碎片，加剧锁争用。
(2)降低IO吞吐。由于swapable huge page，在swap时需要切分成原有的4K的页。Oracle的测试数据显示会降低30%的IO吞吐。
3.测试性能
查看系统版本：
uname -r
Linux ckl-master1 2.6.32-504.el6.x86_64
查看当前的Huge pages设置：
grep Huge /proc/meminfo
AnonHugePages: 5971968 kB
HugePages_Total: 0
HugePages_Free: 0
HugePages_Rsvd: 0
HugePages_Surp: 0
Hugepagesize: 2048 kB
查看是否开启了Huge pages：
cat /sys/kernel/mm/transparent_hugepage/enabled
[always] madvise never
测试THP开启redis性能：
/usr/local/redis/bin/redis-benchmark -P 4 -t set -r 5000000 -n 10000000
====== SET ======
10000000 requests completed in 41.43 seconds
50 parallel clients
3 bytes payload
keep alive: 1
98.62% <= 1 milliseconds
98.62% <= 2 milliseconds
98.63% <= 4 milliseconds
98.63% <= 5 milliseconds
98.63% <= 7 milliseconds
98.63% <= 12 milliseconds
98.64% <= 13 milliseconds
98.64% <= 14 milliseconds
98.65% <= 15 milliseconds
98.66% <= 16 milliseconds
98.68% <= 17 milliseconds
98.69% <= 18 milliseconds
98.74% <= 19 milliseconds
98.83% <= 20 milliseconds
98.96% <= 21 milliseconds
99.13% <= 22 milliseconds
99.32% <= 23 milliseconds
99.49% <= 24 milliseconds
99.60% <= 25 milliseconds
99.67% <= 26 milliseconds
99.74% <= 27 milliseconds
99.79% <= 28 milliseconds
99.81% <= 29 milliseconds
99.84% <= 30 milliseconds
99.85% <= 31 milliseconds
99.86% <= 32 milliseconds
99.86% <= 33 milliseconds
99.87% <= 34 milliseconds
99.87% <= 35 milliseconds
99.88% <= 36 milliseconds
99.89% <= 37 milliseconds
99.89% <= 38 milliseconds
99.89% <= 39 milliseconds
99.89% <= 40 milliseconds
99.90% <= 42 milliseconds
99.90% <= 43 milliseconds
99.91% <= 44 milliseconds
99.91% <= 49 milliseconds
99.91% <= 53 milliseconds
99.91% <= 54 milliseconds
99.92% <= 55 milliseconds
99.92% <= 56 milliseconds
99.93% <= 57 milliseconds
99.93% <= 62 milliseconds
99.93% <= 63 milliseconds
99.93% <= 65 milliseconds
99.93% <= 66 milliseconds
99.93% <= 67 milliseconds
99.94% <= 68 milliseconds
99.94% <= 74 milliseconds
99.94% <= 76 milliseconds
99.94% <= 77 milliseconds
99.94% <= 78 milliseconds
99.95% <= 79 milliseconds
99.95% <= 81 milliseconds
99.95% <= 82 milliseconds
99.95% <= 83 milliseconds
99.95% <= 85 milliseconds
99.96% <= 86 milliseconds
99.96% <= 91 milliseconds
99.96% <= 92 milliseconds
99.97% <= 93 milliseconds
99.97% <= 95 milliseconds
99.97% <= 96 milliseconds
99.98% <= 97 milliseconds
99.99% <= 98 milliseconds
99.99% <= 99 milliseconds
99.99% <= 100 milliseconds
99.99% <= 101 milliseconds
99.99% <= 126 milliseconds
100.00% <= 157 milliseconds
100.00% <= 158 milliseconds
100.00% <= 168 milliseconds
100.00% <= 169 milliseconds
100.00% <= 169 milliseconds
241382.62 requests per second
关闭THP：
echo never > /sys/kernel/mm/transparent_hugepage/enabled
cat /sys/kernel/mm/transparent_hugepage/enabled
always madvise [never]
测试关闭THPredis性能:
/usr/local/redis/bin/redis-benchmark -P 4 -t set -r 5000000 -n 10000000
====== SET ======
10000000 requests completed in 26.72 seconds
50 parallel clients
3 bytes payload
keep alive: 1
99.70% <= 1 milliseconds
99.71% <= 2 milliseconds
100.00% <= 3 milliseconds
100.00% <= 3 milliseconds
374321.53 requests per second
关闭THP后，redis的响应时间明显缩短，建议关闭THP,需要重启redis进程