hdf5/test/SWMR_POSIX_Order_UG.txt
2017-01-26 14:34:12 -05:00

95 lines
3.7 KiB
Plaintext

POSIX Write Order Test Instructions
===================================
Purpose
-------
This documents shows the requirments, implementaion design and instructions
of building and running the POSIX Write Order test. The name of the
test is twriteorder and it resides in the test/ directory.
Requirements
------------
The test is to verify that the write order is strictly consistent.
The SWMR feature requires that the order of write is strictly consistent.
"Strict consistency in computer science is the most stringent consistency
model. It says that a read operation has to return the result of the
latest write operation which occurred on that data item."--
(http://en.wikipedia.org/wiki/Linearizability#Definition_of_linearizability).
This is also an alternative form of what POSIX write require that after a
write operation has returned success, all reads issued afterward should
get the same data the write has written.
Implementation Design
---------------------
The test simulates what SWMR does by writing chained blocks and see if
they can be read back correctly.
There is a writer process and a read process.
The file is divided into 2KB partitions. Then writer writes 1 chained
block, each of 1KB big, in each partition after the first partition.
Each chained block has this structure:
Byte 0-3: offset address of its child block. The last child uses 0 as NULL.
Byte 4-1023: some artificial data.
The child block address of Block 1 is NULL (0).
The child block address of Block 2 is the offset address of Block 1.
The child block address of Block n is the offset address of Block n-1.
After all n blocks are written, the offset address of Block n is written
to the offset 0 of the first partition.
Therefore, by the time the offset address of Block n is written to this
position, all n chain-linked blocks have been written.
The other reader processes will try to read the address value at the
offset 0. The value is initially NULL(0). When it changes to non-zero,
it signifies the writer process has written all the chain-link blocks
and they are ready for the reader processes to access.
If the system, in which the writer and reader processes run, the readers
will always get all chain-linked blocks correctly. If the order of write
is not maintained, some reader processes may found unexpect block data.
Building the Tests
------------------
The name of the test is twriteorder in the test directory. It is added
to the test suite and is built during the "make" process and is run by
the test_usecases.sh test. Users may inspect test/test_usecases.sh.in
to see the examples of testing.
Running the Tests
-----------------
twriteorder test accepts the following options:
$ ./twriteorder -h
usage: twriteorder [OPTIONS]
OPTIONS
-h Print a usage message and exit
-l w|r launch writer or reader only. [default: launch both]
-b N Block size [default: 1024]
-p N Partition size [default: 2048]
-n N Number of linked blocks [default: 512]
More Examples
-------------
# run test with default parameters and launch both writer and reader
#processes.
$ twriteorder
# run test with blocksize of 1000 bytes (default is 1024 bytes).
$ twriteorder -b 1000
# run test with partition size of 3000 bytes (default is 2048 bytes).
$ twriteorder -p 3000
# run test with 2000 linked blocks (default is 512 blocks).
$ twriteorder -n 2000
# Launch only the writer process.
$ twriteorder -l w
# Launch only the reader process.
$ twriteorder -l r
Note that if you want to launch the writer and the reader processes
manually (for example in different machines sharing a common file system),
you need to start the writer process (-l w) first, and then the reader
process (-l r).