基于PostGIS的高级应用(1)--PgRouting百万级别路网查询优化
前文《基于PgRouting的GIS网络分析--数据准备》描述了如何进行数据准备工作,入门的朋友可以参考如何搭建环境,导入数据和建立索引等准备。pgrouting是postgis的插件,主要做网络分析等业务使用,一般一个地区,一个城市几万级别的路网,查询是非常快速的,但是全国路网动辄几百万,几千万的路网规模,默认查询就非常的慢了。于是,本文主要以dijkstra算法,安装pg的单机默认配置,重点阐述如何“动脑经”加速路径查询速度,而不是单纯依靠机器配置(毕竟再牛逼的机器也架不住无脑的大量运算啊),当然本文的方法并不是非常规范和标准,但提供了一个解决问题的思路,即大量路网的复杂查询优化一定要避免全表查询,尽量减少计算!!!
一 全表路径分析查询
北京路网.png
查询路网的线数据规模:
network=# select count(*) from ways;
count
---------
1250371
(1 row)
以dijkstra算法示例查询:
规划A_B的路径查询.png
如上图:A点坐标[115.2,39.8],B点坐标[115.4,40],A点的附近对应道路的gid是487371,B点附近对应道路的gid是62553(gid事先查询好的,测试就不写如何获取坐标附近的道路gid),考虑到道路有单行道的关系,所以有通行权重cost和反向权重reverse_cost ,查询语句如下:
SELECT * FROM pgr_dijkstra('SELECT gid as id,snodeid as source,enodeid as target,length::float as cost,rev_length::float as reverse_cost FROM ways
',487371,62553,true);
返回93行记录,平均耗时5.8s,但是如图可知,其实AB两点比较近,而大部分路网其实对他们的计算根本就没关系,于是我们考虑在一开始查询时就规避无效路网。
二 矩形范围过滤
矩形过滤
我们发现,AB之间范围及其附近的路网就足够分析出路径,而大量其他数据是没有任何影响作用的,我们设AB两点构成一个矩形,然后缓冲2km作为备用参与分析道路(即红色斜线部分),语句如下:
SELECT * FROM pgr_dijkstra('SELECT gid as id,snodeid as source,enodeid as target,length::float as cost,rev_length::float as reverse_cost FROM ways
where st_intersects(geom,st_buffer(ST_PolygonFromText(''POLYGON((115.2 39.8,115.4 39.8,115.4 40,115.2 40,115.2 39.8))'',4326),0.02))
',487371,62553,true);
返回93行记录,平均耗时150 ms。
实验结果证明:与全表查询的分析结果一致,但全表查询是矩形查询耗时的 5800/150约49倍,明显优化速度还是很明显的。
矩形问题.png
但是矩形查询也存在一个问题,当AB两点的经度接近(极端情况就是一致),那么两点只能构成一个 水平或者垂直的 线段(无法构造矩形区域),绝大部分形成一个细长的面区域如上图,这种情况下,因不能构造矩形筛选区域,或者说构造的区域过于狭窄无法满足路径查询的要求,采用“线性过滤”会更恰当些。
三 线性范围过滤
线性过滤示意图.png
AB两点坐标接近垂直或水平时,可选用线性查询。举例:AB两点不变,根据AB两点坐标构成线,缓冲5公里(线比矩形那个要大,尽量将可能的道路加入分析),查询语句如下:
SELECT * FROM pgr_dijkstra('SELECT gid as id,snodeid as source,enodeid as target,length::float as cost,rev_length::float as reverse_cost FROM ways
where st_intersects(geom,st_buffer(ST_LineFromText(''LineString(115.2 39.8,115.4 40)'',4326),0.05))
',487371,62553,true);
耗时:180ms,效果仍然比较明显。
四 网格筛选过滤
三四节作者猜想了以矩形和线性做查询筛选(线性更通用,不推荐矩形查询),但是他们只能处理两点比较近的时候,筛选出一小部分区域作分析标本,随着两点朝相反对角线拉大,以上图形构成的查询区域也随之变得很大,即使有索引,但是pg的查询优化器发现查询的数量非常大而不是小部分时,可能不走索引,也就是说,随着两点距离变大,越来越接近于全表查询(甚至比全表查询还慢)。这种情况下,作者采用网格对路段分组,如下图:
城市路网网格化.png
图形可视化,形象生动可见路网和格网的关系,但我们还是客观具体的看下表中数据的关系如下。
查询网格:
network=# \d maps
数据表 "public.maps"
栏位 | 类型 | 修饰词
-------+------------------------+--------
mapid | integer | 非空
geom | geometry(Polygon,4326) |
索引:
"maps_pkey" PRIMARY KEY, btree (mapid)
"maps_geom_index" gist (geom)
network=# select * from maps limit 10;
mapid | geom
--------+----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
595756 | 0103000020E6100000010000000500000065BDBD10976F5D402753A278D7DF434065BDBD10976F5D407D776D6786EA4340C85C19549B775D407D776D6786EA4340C85C19549B775D402753A278D7DF434065BDBD10976F5D402753A278D7DF4340
555571 | 0103000020E61000000100000005000000463150849AC75C4025A93394E69F4240463150849AC75C40AB6B40C694AA42406270E6BA9DCF5C40AB6B40C694AA42406270E6BA9DCF5C4025A93394E69F4240463150849AC75C4025A93394E69F4240
615867 | 0103000020E6100000010000000500000074D6E1C79CB75D40CA83DB771895444074D6E1C79CB75D4024EAEC01C69F44407C61536399BF5D4024EAEC01C69F44407C61536399BF5D40CA83DB771895444074D6E1C79CB75D40CA83DB7718954440
615707 | 0103000020E61000000100000005000000E144B24F99775D4091A8275E2B554440E144B24F99775D403459CC9DD35F4440EE76FF50977F5D403459CC9DD35F4440EE76FF50977F5D4091A8275E2B554440E144B24F99775D4091A8275E2B554440
605752 | 0103000020E610000001000000050000007EEB1E34964F5D40911DA72A253544407EEB1E34964F5D40D20A5FA1873E4440A3C87B5192575D40D20A5FA1873E4440A3C87B5192575D40911DA72A253544407EEB1E34964F5D40911DA72A25354440
555451 | 0103000020E61000000100000005000000A71F798C97875C4093814DE7948A4240A71F798C97875C40BA5D58CC429542406E34FC7E9C8F5C40BA5D58CC429542406E34FC7E9C8F5C4093814DE7948A4240A71F798C97875C4093814DE7948A4240
595632 | 0103000020E610000001000000050000009238F1F69C0F5D40308FCA877FCA43409238F1F69C0F5D40FE13F9812DD5434081CFEE149B175D40FE13F9812DD5434081CFEE149B175D40308FCA877FCA43409238F1F69C0F5D40308FCA877FCA4340
625533 | 0103000020E6100000010000000500000098C4D5D890D75C40E3B5BF7161CA444098C4D5D890D75C40F392BDAD0DD54440F0826F3794DF5C40F392BDAD0DD54440F0826F3794DF5C40E3B5BF7161CA444098C4D5D890D75C40E3B5BF7161CA4440
615633 | 0103000020E610000001000000050000008947A0C997175D406F0490771A7544408947A0C997175D406A69B0A1C27F44407B9CEDFA9A1F5D406A69B0A1C27F44407B9CEDFA9A1F5D406F0490771A7544408947A0C997175D406F0490771A754440
615663 | 0103000020E610000001000000050000007DAF424697175D407346393D149544407DAF424697175D407CCE61E7BB9F44400418F9699A1F5D407CCE61E7BB9F44400418F9699A1F5D407346393D149544407DAF424697175D407346393D14954440
(10 行记录)
查询路网
network=# \d ways
数据表 "public.ways"
栏位 | 类型 | 修饰词
------------+---------------------------+---------------------------------------------
gid | integer | 非空 默认 nextval('ways_gid_seq'::regclass)
name | character varying(128) |
pyname | character varying(128) |
mapid | integer |
id | character varying(13) |
kind_num | character varying(2) |
kind | character varying(30) |
width | character varying(3) |
direction | character varying(1) |
toll | character varying(1) |
const_st | character varying(1) |
undconcrid | character varying(13) |
snodeid | integer |
enodeid | integer |
funcclass | character varying(2) |
detailcity | character varying(1) |
through | character varying(1) |
unthrucrid | character varying(13) |
ownership | character varying(1) |
road_cond | character varying(1) |
special | character varying(1) |
admincodel | character varying(6) |
admincoder | character varying(6) |
uflag | character varying(1) |
onewaycrid | character varying(13) |
accesscrid | character varying(13) |
speedclass | character varying(1) |
lanenums2e | character varying(2) |
lanenume2s | character varying(2) |
lanenum | character varying(1) |
vehcl_type | character varying(32) |
elevated | character varying(1) |
structure | character varying(1) |
usefeecrid | character varying(13) |
usefeetype | character varying(1) |
spdlmts2e | character varying(4) |
spdlmte2s | character varying(4) |
spdsrcs2e | character varying(1) |
spdsrce2s | character varying(1) |
dc_type | character varying(1) |
nopasscrid | character varying(13) |
geom | geometry(LineString,4326) |
length | double precision |
rev_length | double precision |
x1 | double precision |
y1 | double precision |
x2 | double precision |
y2 | double precision |
索引:
"ways_pkey" PRIMARY KEY, btree (gid)
"mapid_index" btree (mapid)
"ways_enodeid_idx" btree (enodeid)
"ways_geom_idx" gist (geom)
"ways_snodeid_idx" btree (snodeid)
network=# select a.gid,a.mapid from ways a,(select mapid from maps limit 1) b where a.mapid=b.mapid limit 10;
gid | mapid
---------+--------
112434 | 595756
112440 | 595756
117555 | 595756
23611 | 595756
1041239 | 595756
1193746 | 595756
694218 | 595756
735844 | 595756
739260 | 595756
740230 | 595756
(10 行记录)
网格和路网之间已经根据mapid建立了关系,路网中mapid建立了索引,格网中mapid是主键。
根据图形之前的关系,我们的思路是:根据AB两点建立直线,对该直线建立一定范围内的缓冲面,缓冲面查询与哪些网格有相交关系(相交意味着这些网格是有效的分析网格,其他网格就没任何关系了),直接把这些有效的网格中的路段,作为分析的样本。
测试范围.png
全表查询AB路径如下:
--测试样本
network=# select count(*) from ways;
count
---------
1250371
(1 行记录)
network=# SELECT * FROM pgr_dijkstra('SELECT gid as id,snodeid as source,enodeid as target,length::float as cost,rev_length::float as reverse_cost FROM ways'
,647331,856772,true);
--耗时8.1s
线性查询AB路径如下:
network=# select count(*) from ways where geom&&st_buffer(ST_LineFromText('LineString(114.53247 37.34692,118.125 39.82983)',4326),0.08);
count
--------
678892
(1 行记录)
network=# SELECT * FROM pgr_dijkstra('SELECT gid as id,snodeid as source,enodeid as target,length::float as cost,rev_length::float as reverse_cost FROM ways
where geom&&st_buffer(ST_LineFromText(''LineString(114.53247 37.34692,118.125 39.82983)'',4326),0.08)
',647331,856772,true);
--耗时4.5s
网格查询AB路径如下:
network=# select count(*) from ways where mapid in ( select mapid from maps where st_intersects(geom,st_buffer(ST_LineFromText('LineStr
ing(114.53247 37.34692,118.125 39.82983)',4326),0.08)));
count
--------
127914
(1 行记录)
network=# SELECT * FROM pgr_dijkstra('SELECT gid as id,snodeid as source,enodeid as target,length::float as cost,rev_length::float as reverse_cost FROM ways
where mapid in (select mapid from maps where st_intersects(geom,st_buffer(ST_LineFromText(''LineString(114.53247 37.34692,118.125 39.82983)'',4326),0.08)))
',647331,856772,true);
--耗时1.8s
表格对比如下:
查询方式 | 查询数据量 | 查询时间 |
---|---|---|
全表查询 | 1250371 | 8.1 s |
线性查询 | 678892 | 4.5 s |
网格查询 | 127914 | 1.8 s |
虽然不是太满意结果,但是结果的确体现了“智慧的结晶”了,还算欣慰。
五 道路属性过滤
以全国路网分析简介(未测试),虽然几百万路网,但是 高速,快速路这种道路并不是很多,在做远距离路径分析时,优先选择起点最近的高速A1,终点附近的高速B1,那么先规划AA1,再A1B1,再B1B,将全国路网全表查询细分成高速路路径分析,高速和城市道路分析,细节部分也可采用范围叠加规划。同样的,对完整路网拆分成高速国道的一级查询,省道县道的二级查询,乡道的三级查询等等,尽可能将路网从整体分析中抽离出去。
五 总结
在近距离路径分析时,可考虑范围进行过滤,然后查询过滤后的路网。在远距离路径分析时,可以实际的条件做过滤。在城市道路时,远距离规划车辆路径,应优先规划起点到绕城,绕城出口到终点。
当然,以上测试都是单机测试,全部思路是减少查询路网数量,是动脑经的结果,没有在数据库的配置,硬件的优化上下功夫。作者将在适当时机,移入pgxl进行规划分析测试,看集群是否有利于查询速度提升。