IOR's Algorithm

DRAFT VERSION

1 Introduction

This is a summary of the functions and the algorithms of the LLNL IOR performance program, originally created by Kim Yates.

2.1 Current IOR Algorithm Design

2.1.1 Environment Variables Used by IOR

Environment variables are used to modify the actions of the IOR program.

REPORT_FILE_NAME

The report file name.

TEST_REPETITIONS

How many times to do outermost test.

TEST_PROCS_MAX

Largest number of porcesses user wants to do I/O.

TEST_PROCS_MIN

Smallest number of processes user wants to do I/O.

TEST_PROCS_DEC

If >0, decrement number of processes used.

TEST_PROCS_DIV

If >1, divide number of processes used.

TEST_READ_ONLY

Test read only, no write.

TEST_WRITE_ONLY

Test write only, no read.

DELETE_TEST_FILE

Delete test file after tests are completed.

TEST_WRITE_OVER

Rewrite over existing file.

IOR_DEBUG

Debugging printfs.

SHOW_STATS

Show times.

USE_MPI_TIMER

Use MPI_Timer instead of gettimeofday.

SHOW_RANDOM_LIST

Show random parameters when built.

SHOW_XFERS

Show the results of transfers.

SHOW_DELAYS

Show when we delay.

DO_COLLECTIVE_XFERS

Do collective transfers.

DO_MULTIBLOCK_XFERS

Do multiblock transfers.

TEST_FILE_PATH

Change default file path.

TEST_FILE_NAME

Change default file name.

PREFIX_FOR_OPEN

Special prefix for openning files. (??)

BLOCK_SIZE_MAX

Largest block to try.

BLOCK_SIZE_MIN

Smallest block to try.

BLOCK_SIZE_DEC

If >0, decrement block size.

BLOCK_SIZE_DIV

If >1, divide block size.

SUBBLOCK_SIZE_MAX

Largest subblock to try.

SUBBLOCK_SIZE_MIN

Smallest subblock to try.

SUBBLOCK_SIZE_DEC

If >0, decrement subblock size.

SUBBLOCK_SIZE_DIV

If >1, divide subblock size.

BUFFER_SIZE

New data write/read buffer size.

FILE_SIZE_MAX_MB

Test file size maximum is a megabyte.

FILE_SIZE_MAX

Define test file size maximum.

FILE_SIZE_MIN_MB

Test file size minimum is a megabyte.

FILE_SIZE_MIN

Define test file size minimum.

FILE_SIZE_DEC_MB

Decrement file size by a megabyte.

FILE_SIZE_DEC

If >0, decrement file size.

FILE_SIZE_DIV_MB

Divide file size by a megabyte.

FILE_SIZE_DIV

If >1, divide file size.

DO_CHECK_READ

Check that the data read is as expected.

DO_CHECK_WRITE

Read the data back immediately and check it.

TEST_ACCESS_PATTERNS

Test various access patterns (Regular, Random/InOrder Random/Random)

TEST_FILECREATE_PATTERN

Per Job or per process.

MIN_INTS_PER_CHUNK

Minimum number of integers per chunk.

MAX_INTS_PER_CHUNK

Maximum number of integers per chunk.

CHUNK_SIZE_SEED

Random seed.

MULTIBLOCK_REGION_SIZE

Total bytes in multiblock transvers will be <= this.

NUM_DATASETS

Number of datasets.

INTERTEST_DELAY_SECS

Delay before open to let previous I/O to finish.

H5_DSET_CHUNK_SIZE

• If !=0, number of elements in a chunk used to grow an extendable dataset
• If <0, use default extension chunk size
• If >0, specifies the chunk size
• If 0 (default), use non-extendable datasets

H5_DSET_EXTEND_SIZE

• If >0, number of elements to grow an extendable dataset
• If <=0 (default), use non-extendable datasets

MPI_INFO_KEYSVALS

MPI info keys/vals.

2.1.2 Steps Taken by IOR Program

• Number of processes can be changed for each loop iteration
• Block size vary for acesses which are "Regular"
• Files can be created per Job or Process
• Process 0 does all time management stuff
• Do Write/Read:
• Open file
• Set file access template
• Align datasets for better performance
• Create memory dataspace
• Set up the transfer properties list
• Create file if writing, open file if reading
• Determine sizes of first (n - 1) datasets and nth dataset
• Loop reading/writing datasets
• If Writing:
• Regular pattern: each process transfers 1 block at a time
• Can do multiblock transfers only if filecreate per job
• Non-regular does random sized chunks
• Writing can create extendable datasets
• Create dataset collectively
• If Reading:
• Open dataset
• Set up file for hyperslab
• If multiblock transfers, set up accordingly
• Otherwise, use 1-chunk hyperslabs
• If Writing:
• Might have to extend dataset collectively
• Select hyperslab
• Set up memory hyperslab
• If Writing:
• Do write (collectively, if need be)
• If Reading:
• Do read (collectively, if need be)
• Check write, if specified
• Check read, if specified
• Set up for next loop iteration
• Collectively close dataset
• After loop:
• Participate in collective extensions which remain
• Participate in collective transfers which remain
• Close file
• If Writing:
• Get stat64() information about file (if we have stat64()) and hope that it's the size we expect
• Report max speeds

Albert Cheng & Bill Wendling
National Center for Supercomputing Applications
Send comments to
hdfparallel@ncsa.uiuc.edu

Last Modified: 14. December, 2001