hdf5/test/SWMR_UseCase_UG.txt

1. Title:
     User Guide for SWMR Use Case Programs

2. Purpose:
     This is a User Guide of the SWMR Use Case programs. It describes the use
    case program and explain how to run them.

2.1. Author and Dates:
     Version 2: By Albert Cheng, 2013/06/18.
     Version 1: By Albert Cheng, 2013/06/01.


%%%%Use Case 1.7%%%%

3. Use Case [1.7]:
     Appending a single chunk

3.1. Program name:
     use_append_chunk

3.2. Description:
     Appending a single chunk of raw data to a dataset along an unlimited
     dimension within a pre-created file and reading the new data back.

     It first creates one 3d dataset using chunked storage, each chunk
     is a (1, chunksize, chunksize) square.  The dataset is (unlimited,
     chunksize, chunksize). Data type is 2 bytes integer.  It starts out
     "empty", i.e., first dimension is 0.

     The writer then appends planes, each of (1,chunksize,chunksize)
     to the dataset. Fills each plan with plane number and then writes
     it at the nth plane. Increases the plane number and repeats till
     the end of dataset, when it reaches chunksize long. End product is
     a chunksize^3 cube.

     The reader is a separated process, running in parallel with
     the writer.  It reads planes from the dataset.  It expects the
     dataset is being changed (growing). It checks the unlimited dimension
     (dimension[0]). When it increases, it will read in the new planes, one
     by one, and verify the data correctness.  (The nth plan should contain
     all "n".) When the unlimited dimension grows to the chunksize (it
     becomes a cube), that is the expected end of data, the reader exits.

3.3. How to run the program:
     Simplest way is
     $ use_append_chunk

     It creates a skeleton dataset (0,256,256) of shorts. Then fork off
     a process, which becomes the reader process to read planes from the
     dataset, while the original process continues as the writer process
     to append planes onto the dataset.

     Other possible options:

     1. -z option: different chunksize. Default is 256.
     $ use_append_chunk -z 1024

     It uses (1,1024,1024) chunks to produce a 1024^3 cube, about 2GB big.


     2. -f filename: different dataset file name
     $ use_append_chunk -f /gpfs/tmp/append_data.h5

     The data file is /gpfs/tmp/append_data.h5. This allows two independent
     processes in separated compute nodes to access the datafile on the
     shared /gpfs file system.


     3. -l option: launch only the reader or writer process.
     $ use_append_chunk -f /gpfs/tmp/append_data.h5 -l w   # in node X
     $ use_append_chunk -f /gpfs/tmp/append_data.h5 -l r   # in node Y

     In node X, launch the writer process, which creates the data file
     and appends to it.
     In node Y, launch the read process to read the data file.

     Note that you need to time the read process to start AFTER the write
     process has created the skeleton data file. Otherwise, the reader
     will encounter errors such as data file not found.

     4. -n option: number of planes to write/read. Default is same as the
     chunk size as specified by option -z.
     $ use_append_chunk -n 1000	# 1000 planes are writtern and read.

     5. -s option: use SWMR file access mode or not. Default is yes.
     $ use_append_chunk -s 0

     It opens the HDF5 data file without the SWMR access mode (0 means
     off). This likely will result in error. This option is provided for
     users to see the effect of the needed SWMR access mode for concurrent
     access.

3.4. Test Shell Script:
     The Use Case program is installed in the test/ directory and is
     compiled as part of the make process. A test script (test_usecases.sh)
     is installed in the same directory to test the use case programs. The
     test script is rather basic and is more for demonstrating how to
     use the program.


%%%%Use Case 1.8%%%%

4. Use Case [1.8]:
     Appending a hyperslab of multiple chunks.

4.1. Program name:
     use_append_mchunks

4.2. Description:
     Appending a hyperslab that spans several chunks of a dataset with
     unlimited dimensions within a pre-created file and reading the new
     data back.

     It first creates one 3d dataset using chunked storage, each chunk is a (1,
     chunksize, chunksize) square.  The dataset is (unlimited, 2*chunksize,
     2*chunksize). Data type is 2 bytes integer. Therefore, each plane
     consists of 4 chunks.  It starts out "empty", i.e., first dimension is 0.

     The writer then appends planes, each of (1,2*chunksize,2*chunksize)
     to the dataset. Fills each plan with plane number and then writes
     it at the nth plane. Increases the plane number and repeats till
     the end of dataset, when it reaches chunksize long. End product is
     a (2*chunksize)^3 cube.

     The reader is a separated process, running in parallel with
     the writer.  It reads planes from the dataset.  It expects the
     dataset is being changed (growing). It checks the unlimited dimension
     (dimension[0]). When it increases, it will read in the new planes, one
     by one, and verify the data correctness.  (The nth plan should contain
     all "n".) When the unlimited dimension grows to the 2*chunksize (it
     becomes a cube), that is the expected end of data, the reader exits.

4.3. How to run the program:
     Simplest way is
     $ use_append_mchunks

     It creates a skeleton dataset (0,512,512) of shorts. Then fork off
     a process, which becomes the reader process to read planes from the
     dataset, while the original process continues as the writer process
     to append planes onto the dataset.

     Other possible options:

     1. -z option: different chunksize. Default is 256.
     $ use_append_mchunks -z 512

     It uses (1,512,512) chunks to produce a 1024^3 cube, about 2GB big.


     2. -f filename: different dataset file name
     $ use_append_mchunks -f /gpfs/tmp/append_data.h5

     The data file is /gpfs/tmp/append_data.h5. This allows two independent
     processes in separated compute nodes to access the datafile on the
     shared /gpfs file system.


     3. -l option: launch only the reader or writer process.
     $ use_append_mchunks -f /gpfs/tmp/append_data.h5 -l w   # in node X
     $ use_append_mchunks -f /gpfs/tmp/append_data.h5 -l r   # in node Y

     In node X, launch the writer process, which creates the data file
     and appends to it.
     In node Y, launch the read process to read the data file.

     Note that you need to time the read process to start AFTER the write
     process has created the skeleton data file. Otherwise, the reader
     will encounter errors such as data file not found.

     4. -n option: number of planes to write/read. Default is same as the
     chunk size as specified by option -z.
     $ use_append_mchunks -n 1000	# 1000 planes are writtern and read.

     5. -s option: use SWMR file access mode or not. Default is yes.
     $ use_append_mchunks -s 0

     It opens the HDF5 data file without the SWMR access mode (0 means
     off). This likely will result in error. This option is provided for
     users to see the effect of the needed SWMR access mode for concurrent
     access.

4.4. Test Shell Script:
     The Use Case program is installed in the test/ directory and is
     compiled as part of the make process. A test script (test_usecases.sh)
     is installed in the same directory to test the use case programs. The
     test script is rather basic and is more for demonstrating how to
     use the program.


%%%%Use Case 1.9%%%%

5. Use Case [1.9]:
     Appending n-1 dimensional planes

5.1. Program names:
     use_append_chunk and use_append_mchunks

5.2. Description:
     Appending n-1 dimensional planes or regions to a chunked dataset where
     the data does not fill the chunk.

     This means the chunks have multiple planes and when a plane is written,
     only one of the planes in each chunk is written. This use case is
     achieved by extending the previous use cases 1.7 and 1.8 by defining the
     chunks to have more than 1 plane.  The -y option is implemented for both
     use_append_chunk and use_append_mchunks.

5.3. How to run the program:
     Simplest way is
     $ use_append_mchunks -y 5

     It creates a skeleton dataset (0,512,512), with storage chunks (5,512,512)
     of shorts. It then proceeds like use case 1.8 by forking off a reader
     process.  The original process continues as the writer process that
     writes 1 plane at a time, updating parts of the chunks involved. The
     reader reads 1 plane at a time, retrieving data from partial chunks.

     The other possible options will work just like the two use cases.

5.4. Test Shell Script:
     Commands are added with -y options to demonstrate how the two use case
     programs can be used as for this use case.