为什么PostgreSQL比MongoDB还快之续篇(WiredTiger引擎)

今年的DTCC大会上，MongoDB中国的唐总带来了《如何在3.0实现7-10倍性能提升》。演讲时顺便倒了点苦水：一些其它数据库喜欢拿MongoDB进行性能PK，但MongoDB之前的开发一直没有怎么关注性能这块，以前也没有发布过官方的性能测试数据，所以结果可想而知。
但是，MongoDB 3.0带来了新的WiredTiger存储引擎,不再像以前(MMAPv1引擎)那样受制于OS内存映射，性能有7-10倍的提升。

这里有一份MongoDB的官方性能测试报告，根据这份测试报告，WiredTiger的性能提升主要表现在数据压缩和并行加载上。


由于WiredTiger不是默认引擎(根据唐总的说法，以后的版本会考虑作为默认引擎) 。我之前的那篇测试，对比的还是MongoDB的老引擎MMAPv1。
http://blog.chinaunix.net/xmlrpc.php?r=blog/article&uid=20726500&id=4960138

下面用同样的测试方法看看WiredTiger引擎的表现(嫌测试过程写得太长的话，可以直接跳到后面看测试总结)。

1. MongoDB WiredTiger引擎的测试

1）启用MongoDB的WiredTiger引擎

-bash-4.1$ mongod --dbpath /data/db2 --storageEngine wiredTiger
2015-04-18T07:51:26.647+0800 I STORAGE  [initandlisten] wiredtiger_open config: create,cache_size=1G,session_max=20000,eviction=(threads_max=4),statistics=(fast),log=(enabled=true,archive=true,path=journal,compressor=snappy),checkpoint=(wait=60,log_size=2GB),statistics_log=(wait=0),
2015-04-18T07:51:27.561+0800 I CONTROL  [initandlisten] MongoDB starting : pid=21287 port=27017 dbpath=/data/db2 64-bit host=hanode1
2015-04-18T07:51:27.563+0800 I CONTROL  [initandlisten] db version v3.0.2
2015-04-18T07:51:27.563+0800 I CONTROL  [initandlisten] git version: 6201872043ecbbc0a4cc169b5482dcf385fc464f
2015-04-18T07:51:27.564+0800 I CONTROL  [initandlisten] OpenSSL version: OpenSSL 1.0.1e-fips 11 Feb 2013
2015-04-18T07:51:27.564+0800 I CONTROL  [initandlisten] build info: Linux ip-10-171-120-213 2.6.32-220.el6.x86_64 #1 SMP Wed Nov 9 08:03:13 EST 2011 x86_64 BOOST_LIB_VERSION=1_49
2015-04-18T07:51:27.564+0800 I CONTROL  [initandlisten] allocator: tcmalloc
2015-04-18T07:51:27.564+0800 I CONTROL  [initandlisten] options: { storage: { dbPath: "/data/db2", engine: "wiredTiger" } }
2015-04-18T07:51:27.595+0800 I NETWORK  [initandlisten] waiting for connections on port 27017

2）加载数据

-bash-4.1$ time -p mongoimport --type json --collection json_tables --db benchmark /dev/null 2>/dev/null
real 8.53
user 7.99
sys 4.14

加载期间的系统负载

[root@hanode1 ~]# top
top - 08:19:09 up 7 days, 11:19,  5 users,  load average: 0.48, 0.13, 0.04
Tasks: 151 total,   1 running, 150 sleeping,   0 stopped,   0 zombie
Cpu(s): 24.9%us,  9.3%sy,  0.0%ni, 56.8%id,  6.2%wa,  0.5%hi,  2.4%si,  0.0%st
Mem:   1019320k total,   949576k used,    69744k free,    82976k buffers
Swap:  2064376k total,    62092k used,  2002284k free,   215228k cached

PID USER      PR  NI  VIRT  RES  SHR S %CPU %MEM    TIME+  COMMAND
24190 postgres  20   0  549m 219m 3312 S 114.7 22.1   0:08.94 mongoimport
21287 postgres  20   0  634m 331m 4748 S 34.6 33.3   0:44.48 mongod

值得注意的是系统的瓶颈在mongoimport进程(mongoimport的CPU利用率超过了100%)，而不是mongod进程，mongod还有很多的余力。

 3）建索引

-bash-4.1$ echo "db.json_tables.ensureIndex( { \"name\": 1})" |time -p mongo benchmark >/dev/null
real 1.25
user 0.03
sys 0.03
-bash-4.1$ echo "db.json_tables.ensureIndex( { \"type\": 1})" |time -p mongo benchmark >/dev/null
real 0.45
user 0.02
sys 0.02
-bash-4.1$ echo "db.json_tables.ensureIndex( { \"brand\": 1})" |time -p mongo benchmark >/dev/null
real 0.37
user 0.05
sys 0.03

4）查看存储空间大小

-bash-4.1$ mongo benchmark
MongoDB shell version: 3.0.2
connecting to: benchmark
> db.json_tables.stats()
{
    "ns" : "benchmark.json_tables",
    "count" : 100001,
    "size" : 266284846,
    "avgObjSize" : 2662,
    "storageSize" : 43167744,
    "capped" : false,
    "wiredTiger" : {
        "metadata" : {
            "formatVersion" : 1
        },
        "creationString" : "allocation_size=4KB,app_metadata=(formatVersion=1),block_allocation=best,block_compressor=snappy,cache_resident=0,checkpoint=(WiredTigerCheckpoint.2=(addr=\"01e208e781e4d41e9d38e208e881e4369503e3e208e981e446480254808080e402928fc0e40291cfc0\",order=2,time=1429314987,size=43118592,write_gen=9063)),checkpoint_lsn=(2,40610048),checksum=uncompressed,collator=,columns=,dictionary=0,format=btree,huffman_key=,huffman_value=,id=5,internal_item_max=0,internal_key_max=0,internal_key_truncate=,internal_page_max=4KB,key_format=q,key_gap=10,leaf_item_max=0,leaf_key_max=0,leaf_page_max=32KB,leaf_value_max=1MB,memory_page_max=10m,os_cache_dirty_max=0,os_cache_max=0,prefix_compression=0,prefix_compression_min=4,split_deepen_min_child=0,split_deepen_per_child=0,split_pct=90,value_format=u,version=(major=1,minor=1)",
        "type" : "file",
        "uri" : "statistics:table:collection-2--333628209475642491",
        "LSM" : {
            "bloom filters in the LSM tree" : 0,
            "bloom filter false positives" : 0,
            "bloom filter hits" : 0,
            "bloom filter misses" : 0,
            "bloom filter pages evicted from cache" : 0,
            "bloom filter pages read into cache" : 0,
            "total size of bloom filters" : 0,
            "sleep for LSM checkpoint throttle" : 0,
            "chunks in the LSM tree" : 0,
            "highest merge generation in the LSM tree" : 0,
            "queries that could have benefited from a Bloom filter that did not exist" : 0,
            "sleep for LSM merge throttle" : 0
        },
        "block-manager" : {
            "file allocation unit size" : 4096,
            "blocks allocated" : 9066,
            "checkpoint size" : 43118592,
            "allocations requiring file extension" : 9046,
            "blocks freed" : 27,
            "file magic number" : 120897,
            "file major version number" : 1,
            "minor version number" : 0,
            "file bytes available for reuse" : 40960,
            "file size in bytes" : 43167744
        },
        "btree" : {
            "btree checkpoint generation" : 8,
            "column-store variable-size deleted values" : 0,
            "column-store fixed-size leaf pages" : 0,
            "column-store internal pages" : 0,
            "column-store variable-size leaf pages" : 0,
            "pages rewritten by compaction" : 0,
            "number of key/value pairs" : 0,
            "fixed-record size" : 0,
            "maximum tree depth" : 3,
            "maximum internal page key size" : 368,
            "maximum internal page size" : 4096,
            "maximum leaf page key size" : 3276,
            "maximum leaf page size" : 32768,
            "maximum leaf page value size" : 1048576,
            "overflow pages" : 0,
            "row-store internal pages" : 0,
            "row-store leaf pages" : 0
        },
        "cache" : {
            "bytes read into cache" : 267348602,
            "bytes written from cache" : 267938187,
            "checkpoint blocked page eviction" : 0,
            "unmodified pages evicted" : 0,
            "page split during eviction deepened the tree" : 0,
            "modified pages evicted" : 26,
            "data source pages selected for eviction unable to be evicted" : 3,
            "hazard pointer blocked page eviction" : 3,
            "internal pages evicted" : 0,
            "pages split during eviction" : 26,
            "in-memory page splits" : 6,
            "overflow values cached in memory" : 0,
            "pages read into cache" : 9005,
            "overflow pages read into cache" : 0,
            "pages written from cache" : 9063
        },
        "compression" : {
            "raw compression call failed, no additional data available" : 0,
            "raw compression call failed, additional data available" : 0,
            "raw compression call succeeded" : 0,
            "compressed pages read" : 9004,
            "compressed pages written" : 9020,
            "page written failed to compress" : 0,
            "page written was too small to compress" : 43
        },
        "cursor" : {
            "create calls" : 14,
            "insert calls" : 100001,
            "bulk-loaded cursor-insert calls" : 0,
            "cursor-insert key and value bytes inserted" : 266602485,
            "next calls" : 300006,
            "prev calls" : 1,
            "remove calls" : 0,
            "cursor-remove key bytes removed" : 0,
            "reset calls" : 100005,
            "search calls" : 0,
            "search near calls" : 0,
            "update calls" : 0,
            "cursor-update value bytes updated" : 0
        },
        "reconciliation" : {
            "dictionary matches" : 0,
            "internal page multi-block writes" : 2,
            "leaf page multi-block writes" : 28,
            "maximum blocks required for a page" : 345,
            "internal-page overflow keys" : 0,
            "leaf-page overflow keys" : 0,
            "overflow values written" : 0,
            "pages deleted" : 0,
            "page checksum matches" : 204,
            "page reconciliation calls" : 32,
            "page reconciliation calls for eviction" : 26,
            "leaf page key bytes discarded using prefix compression" : 0,
            "internal page key bytes discarded using suffix compression" : 9230
        },
        "session" : {
            "object compaction" : 0,
            "open cursor count" : 14
        },
        "transaction" : {
            "update conflicts" : 0
        }
    },
    "nindexes" : 4,
    "totalIndexSize" : 2826240,
    "indexSizes" : {
        "_id_" : 864256,
        "name_1" : 868352,
        "type_1" : 630784,
        "brand_1" : 462848
    },
    "ok" : 1
}

WiredTiger引擎的压缩效果确实不错，把253MB的原始数据压缩到了41MB。

5）数据查询
匹配9091条记录的查询：
点击(此处)折叠或打开

-bash-4.1$ echo "db.json_tables.find({ brand: 'ACME'}).count()"|mongo benchmark
MongoDB shell version: 3.0.2
connecting to: benchmark
9091
bye
-bash-4.1$ echo "DBQuery.shellBatchSize = 10000000000;db.json_tables.find({ brand: 'ACME'})"|time -p mongo benchmark >/dev/null
real 3.93
user 3.56
sys 0.11

*）第1次测的时间是12秒，有IO等待的时间，上面是第2次测试的结果，即数据已经被缓存了。

看看top的资源占用。
点击(此处)折叠或打开

Tasks: 158 total,   2 running, 156 sleeping,   0 stopped,   0 zombie
Cpu(s): 22.6%us,  0.4%sy,  0.0%ni, 76.8%id,  0.0%wa,  0.0%hi,  0.2%si,  0.0%st
Mem:   1019320k total,   793584k used,   225736k free,    61424k buffers
Swap:  2064376k total,    64096k used,  2000280k free,   236052k cached

PID USER      PR  NI  VIRT  RES  SHR S %CPU %MEM    TIME+  COMMAND
27969 postgres  20   0  752m  67m 9304 R 91.4 6.8   0:03.46 mongo
26391 postgres  20   0  610m 325m  11m S 1.0 32.7   0:10.95 mongod

测试结果和MMAPv1引擎差不多，CPU把时间都耗在客户端的mongo进程上。因为客户端在处理大量输出结果时消耗了太多的资源。

为了排除大量数据处理的误导，下面执行一下0匹配的查询。

-bash-4.1$ echo "DBQuery.shellBatchSize = 10000000000;db.json_tables.find({ brand: 'ACME111'})"|time -p mongo benchmark >/dev/null
real 0.07
user 0.04
sys 0.01

测试结果和MMAPv1引擎也差不多。

再试试0匹配的全表扫描。

-bash-4.1$ echo "db.json_tables.dropIndexes()"|mongo benchmark
MongoDB shell version: 3.0.2
connecting to: benchmark
{
    "nIndexesWas" : 4,
    "msg" : "non-_id indexes dropped for collection",
    "ok" : 1
}
bye
-bash-4.1$ echo "DBQuery.shellBatchSize = 10000000000;db.json_tables.find({ brand: 'ACME111'})"|time -p mongo benchmark >/dev/null
real 0.14
user 0.02
sys 0.03

测试结果和MMAPv1引擎比快了有50%左右。

6）数据插入
先清数据

-bash-4.1$ echo "db.json_tables.drop()"|mongo benchmark
MongoDB shell version: 3.0.2
connecting to: benchmark
true
bye

插入数据

-bash-4.1$ time mongo benchmark --quiet <sample_mongo_inserts.json >/dev/null

real    1m31.420s
user    0m30.832s
sys    0m50.319s

测试结果也和MMAPv1引擎差不多。
并且由于MongoDB的控制台不允许插入大于4k的文档，最后插入的数据没有10万条。

-bash-4.1$ mongo benchmark
MongoDB shell version: 3.0.2
connecting to: benchmark
> db.json_tables.count()
72728

下面看看插入时的系统负载

[root@hanode1 ~]# top
top - 08:02:23 up 7 days, 11:02,  5 users,  load average: 0.00, 0.02, 0.00
Tasks: 154 total,   2 running, 152 sleeping,   0 stopped,   0 zombie
Cpu(s):  3.3%us,  1.5%sy,  0.0%ni, 76.4%id,  0.0%wa,  0.0%hi, 18.8%si,  0.0%st
Mem:   1019320k total,   888680k used,   130640k free,    83272k buffers
Swap:  2064376k total,    62140k used,  2002236k free,   295144k cached

PID USER      PR  NI  VIRT  RES  SHR S %CPU %MEM    TIME+  COMMAND
22470 postgres  20   0  753m  66m 9224 R 88.8 6.7   0:03.94 mongo
21287 postgres  20   0  634m 344m  11m S 22.3 34.6   0:12.53 mongod

这个测试结果也和MMAPv1引擎惊人的相识。那么WiredTiger引擎的插入性能提升在什么地方呢？
其实WiredTiger引擎在插入上主要提升的是并发性能，MMAPv1引擎在并发时是简单粗暴的锁库模式，严重制约并发性能，CPU核心再多性能也上不去。而我的测试是单并发测试，所以看不出MMAPv1引擎的这个致命缺点。

2. 总结

以单并发时的服务端进程CPU消耗作为衡量指标，PG和WiredTiger的对比总结如下：
1）加载
  WiredTiger的性能是PG的3倍（注）
2）插入
 相差不大，WiredTiger小胜（注）
3）全表扫描（0匹配）
 WiredTiger的性能是PG的4倍
4）单点索引扫描（0匹配）
 PG的性能是WiredTiger的4倍
5）数据大小
 PG的数据大小是WiredTiger的3倍

注意，以上加载和插入的这两个数据以单并发时的服务端进程CPU消耗作为衡量指标的，忽略了MongoDB客户端的高CPU消耗，实际场景中也有可能PG更快。之所以这样对比，是因为：

我们假设好的引擎可以在高并发时可以把负载均衡地分散到所有CPU核心上，对这样的引擎，在CPU成为瓶颈的场景中，单并发时的CPU实际占用时间可以作为一个重要的参考指标。

但是，服务端能不能在高并发时实现多CPU核心上的线性或近似线性的性能Scale UP还需看服务端（PG/MongoDB）的素质和具体使用场景。

最后，即使PG在加载和插入的PK上输了也很正常，因为PG是保障了ACID的，WiredTiger却有丢失数据的风险。
Journal 默认不会即时刷盘，系统宕机会丢失最多100MB Journal数据