1 HDF5 Life cycle of data Boeing September 19, 2006

2 Overview “Life cycle” of HDF5 data I/O operations for datasets with different storage layouts Compact dataset Contiguous dataset Datatype conversion Partial I/O for contiguous dataset Chunked dataset I/O for chunked dataset Variable length datasets and I/O

3 Life cycle: what does happen to data when it is transferred from application buffer to HDF5 file? File or other “storage” Virtual file I/O Library internals Object API Application Data buffer H5Dwrite Magic box Unbuffered I/O Data in a file

4 “Life cycle” of HDF5 data: inside the magic box Operations on data inside the magic box Datatype conversion Scattering - gathering Data transformation (filters, compression) Copying to/from internal buffers Concepts involved HDF5 metadata, metadata cache Chunking, chunk cache Data structures used B-trees (groups, dataset chunks) Hash tables Local and Global heaps (variable length data: link names, strings, etc.)

5 “Life cycle” of HDF5 data: inside the magic box Understanding of what is happening to data inside the magic box will help to write efficient applications HDF5 library has mechanisms to control behavior inside the magic box Goals of this and the next talk are to Introduce the basic concepts and internal data structures and explain how they affect performance and storage sizes Give some “recipes” for how to improve performance

6 Operations on data inside the magic box Datatype conversion Examples: float  integer LE  BE 64-bit integer to 16-bit integer (overflow may occur!) Scattering - gathering Data is scattered/gathered from/to user’s buffers into internal buffers for datatype conversion and partial I/O Data transformation (filters, compression) Checksum on raw data and metadata (in 1.8.0) Algebraic transform GZIP and SZIP compressions User-defined filters Copying to/from internal buffers

7 “Life cycle” of HDF5 data: inside the magic box HDF5 metadata Information about HDF5 objects used by the library Examples: object headers, B-tree nodes for group, B-Tree nodes for chunks, heaps, super-block, etc. Usually small compared to raw data sizes (KB vs. MB-GB) Metadata cache Space allocated to handle pieces of the HDF5 metadata Allocated by the HDF5 library in application’s memory space Cache behavior affects overall performance Will cover in the next talk

8 “Life cycle” of HDF5 data: inside the magic box Chunking mechanism Chunking – storage layout where a dataset is partitioned in fixed- size multi-dimensional tiles or chunks Used for extendible datasets and datasets with filters applied (checksum, compression) HDF5 library treats each chunk as atomic object Greatly affects performance and file sizes Chunk cache Created for each chunked dataset Default size 1MB

9 HDF5 file structure User block File header info Version #, etc. Root group Symbol Table Group Symbol Table Object Global Heap Local Heap

10 Writing a contiguous dataset of atomic type DataMetadata Dataspace 3 Rank Dim_2 = 5 Dim_1 = 4 Dimensions Time = 32.4 Pressure = 987 Temp = 56 Attributes Chunked Compressed Dim_3 = 7 Storage info IEEE 32-bit float Datatype

11 I/O operations for HDF5 datasets with different storage layouts Storage layouts Compact Contiguous Chunked I/O performance depends on Dataset storage properties Chunking strategy Metadata cache performance Etc.

12 Application memory Writing a compact dataset Dataset header …………. Datatype Dataspace …………. Attribute 1 Attribute 2 Data Metadata cache File Raw data is stored within the dataset header

13 Writing a contiguous dataset with no datatype conversion User buffer (matrix 5x4x7) Dataset header …………. Datatype Dataspace …………. Attribute 1 Attribute 2 ………… Application memory Metadata cache File Dataset headerDataset raw data

14 Writing a contiguous dataset with conversion Dataset header …………. Datatype Dataspace …………. Attribute 1 Attribute 2 ………… Application memory Metadata cache File Dataset headerDataset raw data Conversion buffer 1MB Dataset raw data

15 Sub-setting of contiguous dataset Series of adjacent rows File N Application data in memory Data is contiguous in a file One I/O operation M rows M

16 Sub-setting of contiguous dataset Adjacent, partial rows File N M … Application data in memory Data is scattered in a file in M contiguous blocks Several small I/O operation N elements

17 Sub-setting of contiguous dataset Extreme case: writing a column File N M … Application data in memory Data is scattered in a file in M contiguous blocks Several small I/O operation 1 element

18 Sub-setting of contiguous dataset Data sieve buffer File N M … Application data in memory Data is scattered in a file in M contiguous blocks 1 element Data is gathered in a sieve buffer in memory 64K memcopy

19 Performance tuning for contiguous dataset Datatype conversion Avoid for better performance Use H5Pset_buffer function to customize conversion buffer size Partial I/O Write/read in big contiguous blocks (at least the size of a block on FS) Use H5Pset_sieve_buf_size to improve performance for complex subsetting

20 Possible tuning work Datatype conversion Use of multiple threads for datatype conversion Partial I/O OS vector I/O Asynchronous I/O

21 Writing chunked dataset Dataset is partitioned into fixed-size multi-dimensional chunks of sizes X/4 x Y/2 x Z Dimension sizes X x Y x Z

22 Extending chunked dataset in any dimension Data can be added in any dimensions Compression is applied to each chunk Datatype conversion is applied to each chunk

23 Writing chunked dataset CB A ………….. Each chunk is written as a contiguous blob Chunks may be scattered all over the file Compression is performed when chunk is evicted from the chunk cache Other filters when data goes through filter pipeline (e.g. encryption) ABC C File Chunk cacheChunked dataset Filter pipeline

24 Writing chunked dataset Dataset_1 header ………… Application memory Metadata cache Chunking B-tree nodes Chunk cache Default size is 1MB Size of chunk cache is set for file Each chunked dataset has its own chunk cache Chunk may be too big to fit into cache Memory may grow if application keeps opening datasets Dataset_N header ………… ………

25 Partial I/O for chunked dataset Build list of chunks and loop through the list For each chunk: Bring chunk into memory Map selection in memory to selection in file Gather elements into conversion buffer and perform conversion Scatter elements back to the chunk Perform conversion when chunk is flushed from chunk cache For each element 3 memcopy performed 12 34

26 Partial I/O for chunked dataset 3 Application memory memcopy Application buffer Chunk Elements participated in I/O are gathered into corresponding chunk

27 Partial I/O for chunked dataset 3 Conversion buffer Gather data Scatter data Application memory Chunk cache On eviction from cache chunk is compressed and is written to the file File Chunk

28 Variable length datasets and I/O Examples of variable-length data String A[0] “the first string we want to write” ………………………………… A[N-1] “the N-th string we want to write” Each element is a record of variable-length A[0] (1,1,0,0,0,5,6,7,8,9) length of the first record is 10 A[1] (0,0,110,2005) ……………………….. A[N] (1,2,3,4,5,6,7,8,9,10,11,12,….,M) length of the N+1 record is M

29 Variable length datasets and I/O Variable length description in HDF5 application typedef struct { size_t length; void *p; }hvl_t; Base type can be any HDF5 type H5Tvlen_create(base_type) ~ 20 bytes overhead for each element Raw data cannot be compressed

30 Variable length datasets and I/O Global heap Application buffer Raw data Elements in application buffer point to global heaps where actual data is stored Global heap

31 Writing VL datasets Dataset header ………… Application memory Metadata cache B-tree nodes Chunk cache ……… Conversion buffer Raw data Global heap Chunk cache VL chunked dataset with selected region File Filter pipeline

32 VL chunked dataset in a file File Dataset header Chunking B-tree Dataset chunksRaw data

33 Variable length datasets and I/O Hints Avoid closing/opening a file while writing VL datasets global heap information is lost global heaps may have unused space Avoid writing VL datasets interchangeably data from different datasets will is written to the same heap If maximum length of the record is known, use fixed-length records and compression

34 Example: Boeing time-segment library application Multiple extendible 1-dim arrays of variable-length records Uses HDF5 Packet Table APIs (H5PT) HDF5 features used Chunked storage Chunk cache Compound and VL datatypes Datatype conversion Partial I/O Complexity affects performance Performance tuning is needed

35 Thank you! Questions ?