Lifecycle based file system I/O evaluation tools.
lcio is designed to synthetically age and test large
parallel filesystems (Luster, GPFS, Ceph, etc)
Cmake > 3.0
MPI > 3.0 (tested with Open MPI 1.10 and 3.1, 1.10 has some issues that were fixed in later versions)
C11 compatible compiler
lcio takes a .ini file that specifies how the run is to be structured, and another .ini file
that defines the size distribution of the files. It supports
two different 'types' of jobs 'file_tree' and 'rw'. 'file_tree' is the synthetic aging
procedure, and 'rw' is a simple IOR type benchmark.
This is to allow for multiple stages in the whole run, stage one being an initial
benchmark, stage two aging the system for some amount of operations and utilization, then stage 3 runs another
benchmark to see how the performance has changed, then another aging step, etc.
The example config file contains some example configurations of both types of runs. The example dist file has an example file size distribution.
A 'size type' looks like [4k, 128M, 23G, 4t, etc] (upper or lower case k,m,g,t).
All 'int' fields should be positive. Putting a negative value is UB.
Aging is accomplished by the 'file_tree' job type. Required fields are:
type = "file_tree"
tmp_dir = directory to write into
engine = one of "posix" or "mpiio"
mpi_num_pes = int
num_files = int, total number of files, split evenly over PEs
buffer_size = size type, size of buffer used to fill a file (i.e. 512k buffer, 2M file => 4 writes)
fsync = 0 | 1 (false, true), call 'fsync' after each operation (write, read, etc)
ops = int, number of aging ops (1 op = delete + write)
overlap = int , number of processes to share a directory
epochs = int, number of times to do the set of aging ops, barrier sync'd
clean = 0 | 1, clean 'tmp_dir' after job is finished
'rw' is a simple IOR type benchmark measuring time and bandwidth. Required fields
are:
type = 'rw'
mode = 'U' | 'S', (unique, shared) process per dir or all processes in same dir
tmp_dir = directory to write into
buffer_size = size type, size of buffer used to fill files of size 'block_size'
engine = 'posix' or 'mpiio'
mpi_num_pes = int
num_files = int, total number of files, split evenly over PEs
block_size = size type, size of files (Buffer_size MUST evenly divide block_size, best to stick to powers of 2)
fsync = 0 | 1 (false, true), call 'fsync' after each operation (write, read, etc)
clean = 0 | 1, clean 'tmp_dir' after job is finished
The distribution ini file has only one section, [dist], with entries like
4k = 100
32k = 45
64k = 32
These are relative counts, so the left hand side is a 'size type' and the right hand side is an 'int'