RebenchR is an advanced disk IO benchmarking toolkit. This toolkit is able to simulate various workload patterns, support five different APIs with the full control of call parameters, collect detailed statistics and visualize it using R.

Usage:

./rebench [OPTIONS] DEVICE

Alternatively, if no arguments are passed on the command line, ./rebench will read
standard input and read its arguments from there. Multiple lines may be used
to specify multiple concurrent workloads.

Arguments:

DEVICE - device or file name to perform operations on.

Options:

-d, --duration
            Duration of the benchmark in seconds.
            Duration can be specified in KB, MB, and GB (add 'k', 'm', or 'g'
            to the end of the number. The benchmark will stop when that much
            data has been processed. If the number is followed by '%',
            the size will be that percentage of the device. If duration is
            set to 'i', rebench will run commands interactively (useful for
            debugging with btrace, and only available in single-threaded mode).
-c, --threads
            Numbers of threads used to run the benchmark.
-b, --block_size
            Size of blocks in bytes to use for IO operations.
            Block size can also be specified in units other than bytes by appending
            'k', 'm', 'g', or '%'.
-s, --stride
            Size of stride in bytes (only applies to sequential workloads).
            Stride can also be specified in units other than bytes by appending
            'k', 'm', 'g', or '%'.
-w, --workload
            Description of a workload to perform.
            Valid options are 'rnd' for a random load, and 'seq' for a sequential load.
-t, --type
            Type of IO calls to use.
            Valid options are 'stateful' for read/write IO,
            'stateless' for pread/pwrite type of IO, 'mmap' for
            memory mapping, 'paio' for POSIX asynchronous IO,
            and 'naio' for native OS asynchronous IO.
-q, --queue-depth
            The number of simultaneous AIO calls.
            Valid only during 'paio', and 'naio' type of runs.
--eventfd
            Use eventfd for aio completion notification.
            Valid only during 'naio' type of runs. Useful for measuring eventfd overhead.
-r, --direction
            Direction in which the operations are performed.
            Valid options are 'formward' and 'backward'.
            This option is only applicable to sequential workloads.
-o, --operation
            The operation to be performed.
            Valid options are 'read', 'write', and 'trim'.
            Note that trim is only available on SSD devices
            via a stateful interface.
-p, --paged
            This options turns off direct IO (which is on by default).
-f, --buffered
            This options turns off flushing (flushing is on by default).
            This option is only applicable for write operations.
-m, --do-atime
            This options turns on atime updates (off by default).
-a, --append
            Open the file in append-only mode (off by default).
            This option is only applicable for sequential writes.
-u, --dist
            The distribution used for random workloads (uniform by default).
            Valid options are 'uniform', 'normal', 'pow' (for power law) and const
            (for accessing one block). This option is only applicable for random workloads.
-i, --sigma
            Controls the random distribution.
            For normal distribution, it is the percent of the data one standard deviation away
            from the mean (5 by default). If sigma = 5, roughly 68% of the time only 10%
            of the data will be accessed. This option is only applicable when distribution is
            set to 'normal'.
            For power distribution, it is the percentage of the size of the device that will
            be the mean accessed location (5 by default).
-l, --local-fd
            Use a thread local file descriptor (by default file descriptors
            are shared across threads).
-n, --silent
            Non-interactive mode. Won't ask for write confirmation, and will
            print machine readable output in the following format:
            [ops per second] [MB/sec] [min ops per second] [max ops per second] [standard deviation]
-j, --offset
            The offset in the file to start operations from.
            By default, this value is set to zero.
            Offset can also be specified in units other than bytes by appending 'k', 'm', 'g',
            or '%'.
-e, --length
            The length from the offset in the file to perform operations on.
            By default, this value is set to the length of the file or the device.
            Length can also be specified in units other than bytes by appending 'k', 'm', 'g',
            or '%'.
-g, --sample-step
            The timestep between IOPS report samples (in milliseconds).
            Defaults to 1000ms. If set to zero, reports latency of every operation.
-z, --pause
            The timestep to wait between a completion of an operation and execution
            of the next operation in microseconds. Defaults to zero.
--drop-caches
            Asks the kernel to drop the cache before running the benchmark.
--output
            A file name to write detailed data output to at each sample step.

R Script

describe.R visualizes latency and throughput statistics.

Rscript describe.R <trace1 filename> <trace1 name> <trace2 filename> <trace2 name> ...

Sample Plots