GeoServer on steroids All you wanted to know about how to make GeoServer faster but you never asked (or you did and no one answered)

Transcription

1 GeoServer on steroids All you wanted to know about how to make GeoServer faster but you never asked (or you did and no one answered) Ing. Andrea Aime, GeoSolutions Ing. Simone Giannecchini, GeoSolutions

2 GeoSolutions Founded in Italy in late 2006 Expertise Image Processing, GeoSpatial Data Fusion Java, Java Enterprise, C++, Python JPEG2000, JPIP, Advanced 2D visualization Supporting/Developing FOSS4G projects GeoTools, GeoServer GeoBatch, GeoNetwork Clients Public Agencies Private Companies

3 Preparing raster inputs

4 Raster Data CheckList Objectives Fast extraction of a subset of the data Fast extraction of overviews Check-list Avoid having to open a large number of files per request Avoid parsing of complex structures Avoid on-the-fly reprojection (if possible) Get to know your bottlenecks CPU vs Disk Access Time vs Memory Experiment with Format, compression, different color models, tile size, overviews, configuration (in GeoServer of course)

5 Problematic Formats PNG/JPEG direct serving Bad formats (especially in Java) No tiling (or rarely supported) Chew a lot of memory and CPU for decompression Mitigate with external overviews NetCDF/grib1 and similar formats Complex formats (often with many subdatasets) Often contains un-calibrated data Must usually use multiple dimensions Use ImageMosaic Must usually massage the data before serving e.g. transpose X,Y,

6 Problematic Formats Ascii Grid, GTOPO30, IDRISI and similar formats are bad ASCII formats are bad No internal tiling, no compression, no internal overviews JPEG2000 (with Kakadu) Extensible and rich, not (always) fast Can be difficult to tune for performance (might require specific encoding options) ECW and MrSID Why bother it s proprietary?

7 Choosing Formats and Layouts To remember: GeoTiff is a swiss knife But you don t want to cut a tree with it! Tremendously flexible, good fir for most (not all) use cases BigTiff pushes the GeoTiff limits farther Single File VS Mosaic VS Pyramids Use single GeoTiff when Overviews and Tiling stay within 4GB No additional dimensions Consider BigTiff for very large file (> 4 GB) Support for tiling Support for Overviews Can be inefficient with very large files + small tiling

8 Choosing Formats and Layouts Use ImageMosaic when: A single file gets too big (inefficient seeks, too much metadata to read, etc..) Multiple Dimensions (time, elevation, others..) Avoid mosaics made of many very small files Single granules can be large Use Tiling + Overviews + Compression on granules Use ImagePyramid when: Tremendously large dataset Too many files / too large files Need to serve at all scales Especially low resolution For single granules (< 2Gb) GeoTiff is generally a good fit

9 Choosing Formats and Layouts Examples: Small dataset: single 2GB GeoTiff file Medium dataset: single 40GB BigTiff Large dataset: 400GB mosaic made of 10GB BigTiff files Extra large: 4TB of imagery, built as pyramid of mosaics of BigTiff/GeoTiff files to keep the file count low

10 GeoTiff preparation STEP 0: get to know your data gdalinfo utility is your friend CheckList Missing CRS Add a.prj file Fix with gdal_translate Missing georeferencing Add a World File Fix with gdal_translate Bad Tiling Fix with gdal_translate Missing Overviews Use gdaladdo Compression Use gdal_translate

11 GeoTiff preparation STEP 1: fix and optimize with gdal_translate Inner Tiling gdal_translate -co "TILED=YES" -co "BLOCKXSIZE=512" -co "BLOCKYSIZE=512" in.tif out.tif Check also GeoTiff driver creation options here CRS and GeoReferencing gdal_translate a_srs EPSG:32619 a_ullr in.tif out.tif STEP 2: add overviews with gdal_addo Leverages on tiff support for multipage files and reduced resolution pages gdaladdo -r cubic output.tif Choose the resampling algorithm wisely Chose the tile size and compression wisely (use GDAL_TIFF_OVR_BLOCKSIZE) Consider external overviews

12 GeoTiff preparation

13 GeoTiff preparation Compression Consider when disk speed/space is an issue Control it with gdal_translate and creation options GeoTiff tiles can be compressed LZW/Deflate are good for lossless compression JPEG is good for visually lossless compression From experience Use LZW/Deflate on geophysical data (DEM, acquisitions) USE JPEG visually lossless with Photometric Interpretation to YCbCr for RGB

14 Use Cases: Time, Elevation and other dimensions MetOc data (support for time, elevation) Data with additional indipendent dimensions WorkFlow Split in multiple GeoTiff files Optimize the files individually Use ImageMosaic Use a DBMS for indexing granules Use File Name based property collectors to turn properties into DB rows attributes Filter by time, elevation and other attributes via OGC and CQL filters Check back up slides for more info!

15 Time, Elevation and other dimensions Indexing multiple dimensions with DB support (video here) datastore.properties timeregex.properties stringregex.properties indexer.properties

16 Time, Elevation and other dimensions

17 Proper Mosaic Preparation ImageMosaic stitches single granules together with basic processing Filtered selection Overviews/Decimation on read Over/DownSampling in memory ColorMask (optional) Mosaic/Stitch ColorMask again (optional) Optimize files as if you were serving them individually Keep a balance between number and dimensions of granules

18 Proper Mosaic Configuration STEP 0: Configure Coverage Access (see slide 22) STEP 1: Configure Mosaic Parameters ALLOW_MULTITHREADING Load data from different granules in parallel Needs USE_JAI_IMAGE_READ set to false (Immediate Mode) Use a proper Tile Size In-memory processing, must not be too large Disk tiling should larger If memory is scarce: USE_JAI_IMAGREAD to true USE_MULTITHREADING to false* Otherwise USE_JAI_IMAGREAD to false ALLOW_MULTITHREADING to true

19 Proper Mosaic Configuration Optional (Advanced): Configure Mosaic Parameters Directly Caching Load the index in memory (using JTS SRTree) Super fast granule lookup, good for shapefiles Bad if you have additional dimension to filter on Based on Soft References, controlled via Java switch SoftRefLRUPolicyMSPerMB ExpandToRGB Expand colormapped imagery to RGB in memory Trade performance for quality SuggestedSPI Default ImageIO Decoder class to use Don t touch unless expert

20 Proper Pyramid Preparation Use gdal_retile for creating the pyramid Prepare the list of tiles to be retiled Create the pyramid with GDAL retile (grab a coffee!) Chunks should not be too small (here 2048x2048) Too many files is bad anyway Use internal Tiling for Larger chunks size If the input dataset is huge use the usedirforeachrow option Too many files in a dir is bad practice Make sure the number of level is consistent Too few bad performance at high scale

21 Proper Pyramid Configuration STEP 0: Configure Coverage Access (see slide 22) STEP 1: Configure Pyramid Parameters ImagePyramid relies on ImageMosaic ALLOW_MULTITHREADING Load data from different granules in parallel Needs USE_JAI_IMAGE_READ set to false (Immediate Mode) Use a proper Tile Size In-memory processing, must not be too large Disk tiling should larger If memory is scarce: USE_JAI_IMAGREAD to true USE_MULTITHREADING to false* Otherwise USE_JAI_IMAGREAD to false ALLOW_MULTITHREADING to true

22 Proper Pyramid Configuration Optional (Advanced): Configure Mosaic Parameters Directly Caching Load the index in memory (using JTS SRTree) Super fast granule lookup, good for shapefiles Bad if you have additional dimension to filter on Based on Soft References, controlled via Java switch SoftRefLRUPolicyMSPerMB ExpandToRGB Expand colormapped imagery to RGB in memory Trade performance for quality SuggestedSPI Default ImageIO Decoder class to use Don t touch unless expert

23 Proper GDAL Formats Configuration Fix Missing/Improper CRS with PRJ or coverage config Fix Missing GeoReferencing with World File Make sure GDAL_DATA is properly configured Use a proper Tile Size In-memory processing, must not be too large Fundamental for striped data! JNI overhead Disk tiling should larger If memory is scarce: USE_JAI_IMAGREAD to true USE_MULTITHREADING to true* Otherwise USE_JAI_IMAGREAD to false USE_MULTITHREADING is ignored

24 Proper JPEG2000 Kakadu Configuration Fix Missing/Improper CRS with PRJ or coverage config Fix Missing GeoReferencing with World File Make sure Kakadu dll/so is properly loaded Use a proper Tile Size In-memory processing Must not be too large Disk tiling should larger If memory is scarce: USE_JAI_IMAGREAD to true USE_MULTITHREADING to true* Otherwise USE_JAI_IMAGREAD to false USE_MULTITHREADING is ignored

25 Proper GeoServer Coverage Options Configuration Make sure native JAI and Image is installed Enable ImageIO native acceleration Enable JAI Mosaicking native acceleration Give JAI enough memory Don t raise JAI memory Threshold too high Rule of thumb: use 2 X #Core Tile Threads (check next slide) Enable Tile Recycling only on trunk Enable Tile Recycling if memory is not a problem

26 Proper GeoServer Coverage Options Configuration Multithreaded Granule Loading Allows to fine tuning multithreading for ImageMosaic Orthogonal to JAI Tile Threads Rule of Thumb: use 2 X #Core Tile Threads Perform testing to fine tune depending on layer configuration as well as on typical requests ImageIO Cache threshold decide when we switch to disk cache (very large WCS requests)

27 Reprojection Performance Vs Quality GeoServer 2.1.x reprojects raster data using a piecewiselinear algorithm The area is divided in rectangular blocks, each having its own affine transform The transformation between the full trigonometric expressions and the linear ones is driven by a tolerance, default value is Larger value will make reprojection faster, but lower the quality -Dorg.geotools.referencing.resampleTolerance=0.5

28 Preparing vector inputs

29 Vector data checklikst What do we want from vector data: Binary data No complex parsing of data structures Fast extraction of a geographic subset Fast filtering on the most commonly used attributes

30 Choosing a format Slow formats WFS GML DXF Good formats, local and indexable Shapefile Directory of shapefiles SDE Spatial databases: PostGIS, Oracle Spatial, DB2, MySQL*, SQL server*

31 Shapefiles vs DBMS Speed comparison vs spatial extent depicted: Shapefile very fast when rendering the full dataset Database faster when extracting a small subset of a very large data set Shapefile no attribute indexing, avoid if filtering on attribute is important (filtering == reading less data, not applying symbols) Database Rich support for complex native filters Use connection pooling (preferably via JNDI) Validate connections (with proper pooling)

32 Shapefile preparation Remove.qix file if present, let GeoServer 2.1.x rebuild it (more efficient) If there are large DBF attributes that are not in use, get rid of them using ogr2ogr, e.g.: ogr2ogr -select FULLNAME,MTFCC arealm.shp tl_2010_08013_arealm.shp If on Linux, enable memory mapping, faster, more scalable (but will kill Windows):

33 Shapefile filtering Stuck with shapefiles and have scale dependent rules like the following? Show highways first Show all streets when zoomed in Use ogr2ogr to build two shapefiles, one with just the highways, one with everything, and build two layers, e.g.: ogr2ogr -sql "SELECT * FROM tl_2010_08013_roads WHERE MTFCC in ('S1100', 'S1200')" primaryroads.shp tl_2010_08013_roads.shp

34 PostGIS specific hints PostgreSQL out of the box configured for very small hardware: Make sure to run ANALYZE after data imports (updates optimizer stats) As usual, avoid large joins in SQL views, consider materialized views If the dataset is massive, CLUSTER on the spatial index: 1.3/ch05.html Careful with prepared statements (bad performance)

35 Optimize styling

36 Use scale dependencies Never show too much data the map should be readable, not a graphic blob. Rule of thumb: 1000 features max in the display

37 Labeling Labeling conflict resolution is expensive, limit to the most inner zooms Halo is important for readability, but adds significant overhead Careful with maxdisplacement, makes for various label location attempts

38 FeatureTypeStyle GeoServer uses SLD FeatureTypeStyle objects as Z layers for painting Each one allocates its own rendering surface (which can use a lot of memory), use as few as possible

39 Use translucency sparingly Translucent display is expensive, use it sparingly

40 Scale dependent rules Too often forgotten or little used, yet very important: Hide layers when too zoomed in (raster/vector example) Progressively show details Add more expensive rendering when there are less features Key to any high performance / good looking map

41 Example

42 Hide as you zoom in Add a MinScaleDenominator to the rule This will make the layer disappear at 1:75000 (towards 1:1)

43 Alternative rendering Simple rendering at low scale (up to 1:2000) More complex rendering when zoomed in (1:1999 and above)

44 Alternative rendering

45 Point symbols 600 loc for 6 different points types Painful

46 Prepare data alter table pointlm add column image varchar; update pointlm set image = 'shop_supermarket.p.16.png' where MTFCC = 'C3081' and (FULLNAME like '%Shopping%' or FULLNAME like '%Mall%'); update pointlm set image = 'peak.png' where MTFCC = 'C3022' update pointlm set image = 'amenity_prison.p.20.png' where MTFCC = 'K1236'; update pointlm set image = 'museum.p.16.png' where MTFCC = 'K2165'; update pointlm set image = 'airport.p.16.png' where MTFCC = 'K2451'; update pointlm set image = 'school.png' where MTFCC = 'K2543'; update pointlm set image = 'christian3.p.14.png' where MTFCC = 'K2582'; update pointlm set image = 'gate2.png' where MTFCC = 'K3066';

47 Dynamic symbolizers

48 Output tuning

49 WMS output formats JPEG PNG 8bit PNG 24bit 23.8KB 66KB 169.4KB 27KB 27KB Compression artifacts Color reduction Large size 64KB

50 WFS output formats Dimension MB HTTP GZip compression is transparent in GeoServer, make sure proxies keep it (or pay 10x price)

51 Tile caching

52 Tile caching with GeoWebCache Tile oriented maps, fixed zoom levels and fixed grid Useful for stable layers, backgrounds Protocols: WMTS, TMS, WMS-C, Google Maps/Earth, VE Speedup compared to dynamic WMS: 10 to 100 times, assuming tiles are already cached (whole layer preseeded) Suitable for: Mostly static layer No (or few) dynamic parameters (CQL filters, SLD params, SQL query params, time/elevation, format options)

53 Embedded GWC advantage No double encoding when using meta-tiling, faster seeding

54 Space considerations Seeding Colorado, assuming 8 cores, one layer, 0.1 sec 756x756 metatile, 15KB for each tile Do yours: Not enough disk space? Set a disk quota Zoom Time to seed Time to seed Tile count Size (MB) level (hours) (days) 13 58, , , ,713, ,855, ,396, ,584,280 3, ,273,037 14,

55 Resource control

56 WMS request limits Max memory per request: avoid large requests, allows to size the server memory (max concurrent request * max memory) Max time per request: avoid requests taking too much time (e.g., using a custom style provided with dynamic SLD in the request) Max errors: best effort renderer, but handling errors takes time

57 WFS request limits Max feature returned, configured as a global limit Return feature bbox: reduce amount of generated GML Per layer max feature count

58 WCS request limits

59 Control flow Control how many requests are executed in parallel, queue others: Increase throughput Control memory usage Enforce fairness More info here

60 Control flow 17% $GEOSERVER_DATA_DIR/controlflow.properties # don't allow more than 16 GetMap requests in parallel ows.wms.getmap=16

61 Auditing Log each and every request Log contents driven by customizable template Summarize and analyze requests with offline tools More info here

62 JVM and deploy configuration

63 Premise The options discussed here are not going to help visibly if you did not prepare the data and the styles They are finishing touches that can get performance up once the major data bottlenecks have been dealt with Check Running in production instructions here

64 JVM settings --server: enables the server JIT compiler --Xms2048m -Xmx2048m: sets the JVM use two gigabytes of memory --XX:+UseParallelOldGC -XX:+UserParallelGC: enables multi-threaded garbage collections, useful if you have more than two cores --XX:NewRatio=2: informs the JVM there will be a high number of short lived objects --XX:+AggressiveOpt: enable experimental optimizations that will be defaults in future versions of the JVM

65 Native JAI and JDK Install native JAI and use a recent Sun JDK! Benchmark over a small data set (the effect is not as visible on larger ones)

66 Setup a local cluster Java2D locks when drawing antialiased vectors Limits scalability severely Use Apache mod_proxy_balance and setup a GeoServer each 2/4 cores mod_proxy_balance GeoServer GeoServer GeoServer

67 Clustering advantage FOSS4G 2010 vector benchmarks (roads/buildings/isolines and so on, over the entire Spain) GeoServer was benchmarked without local clustering 66%

68 Benchmarking

69 Using JMeter Good benchmarking tool Allows to setup multiple thread groups, different parallelelism and request count, to ramp up the load Can use CSV files to generate semi-randomized requests Reports results in a simple table

70 Using JMeter Thread group: how many threads Loop: how many requests HTTP sampler: the request CSV: read request params from CSV Summary table

71 Generating the CSV Simple randomized generation tool built during WMS shootouts, wms_request.py Generate csv with the bbox and width/height to be used in JMeter scripts:./wms_request.py -count region minres maxres 0.1 -minsize maxsize Get it here along with a corresponding JMeter script:

72 Checking results Results table Run the benchmarks 2-3 times, let the results stabilize Save the results, check other optimizations, compare the results

73 Real world deploy

74 Deploy configuration

75 Raster data Whole Italy at 50cm per pixel Over 4TB, updated fully every 3 years (old data still available for historical access) Custom pyramid 100 m per pixel: one image 20m per pixel: mosaic of 20 tiles 4m per pixel: mosaic of few hundred tiles 0.5m per pixel: 9000 tiles Each tile is 10000x10000, with overviews

76 Vector data Cadastral data for the whole Italy, with full history (interval of validity for each parcel) 100 million polygons A query extracts a subset relative to a certain time interval and area the user is allowed to see No data from this table is ever shown below 1:50000 (SLD scale dependencies) Physical table level partitioning (Oracle style) of the table based on geographic area to parallelize and cluster data loading, plus spatial indexing and indexes on commonly filtered upon attributes

77 The End Questions?