NAME

SYNOPSIS

mst-bench [--help] [--trials N] [--file-size GiB]

OPTIONS

--help

Print a basic help message

--trials N

Run N trials and report the average. The default is 3.

--file-size GiB

Use a file of GiB gibibytes for the disk tests. The default is 2 * physical memory size, which is generally enough to overwhelm disk buffering and produce a reasonable estimate of sustainable throughput. However, systems with low physical memory may benefit from large disk cache on the controller which may result in an overestimate of disk speed. In those cases, specifying a larger file system may produce more realistic results.

PURPOSE

The goal is to provide a theoretical maximum to which real programs can be compared. For example, if MST-Bench reports 10 seconds to sequentially read the generated file, then we know that 10 seconds is the theoretical best time that any program can achieve while streaming this file. If fgrep string bench.tmpfile takes 15 seconds, then it is achieving about 2/3 of the maximum theoretical disk throughput. We might then investigate whether fgrep's CPU bottleneck can be reduced to the point where it can keep up with the disk input. Once fgrep is running clone to 10 seconds for this file, we know that it cannot be sped up much further.

DESCRIPTION

It runs the following speed tests:

Sequential memory access for a small array that should result in a high cache hit ratio, where most accesses are satisfied by the cache.

Sequential memory access for a large array that should result in a low cache hit ratio, where many accesses are not satisfied by the cache.

Sequential disk write of a file much larger than physical memory, so that disk buffering has a minimal impact and the reported throughput represents sustainable speed for the disk hardware.

Sequential disk read of a file much larger than physical memory, so that disk buffering has a minimal impact and the reported throughput represents sustainable speed for the disk hardware.

Sequential disk rewrite of a file much larger than physical memory, so that disk buffering has a minimal impact and the reported throughput represents sustainable speed for the disk hardware. In many file systems, overwriting shows different performance characteristics than new writes.

Random disk read of a file much larger than physical memory, so that disk buffering has a minimal impact and the reported throughput represents sustainable speed for the disk hardware. The random read reads the same file as the sequential read, reading every block in the file exactly once but in random order. This provides some idea about the latency of disk access. Note that dividing the file into more random reads of smaller blocks will result in lower performance.

FILES

bench.tmpfile - file generated in current directory for disk speed test

BUGS

AUTHOR

J. Bacon