iops

The complexity of the iops command at present means that GUI support is not (yet) available. You must use it via the command line.

The iops command is an extremely powerful tool that can be used to test a wide variety of scenarios. By default, it runs a medium size random test good for a quick look at a hard drive or solid state drive.

Possibilities include:

Mac OS X caching

DiskTester always tests using user-level (non kernel) public file system APIs . This is important; driver-level tests would not only be dangerous but misleading: no real program writes directly to disk; all real programs use file system APIs.

Unless caching is enabled explicitly with the --caching flag, DiskTester always bypasses the Mac OS X unified buffer cache, so that the true speed of the drive is measured. The Mac OS X UBC is very efficient at what it does, so enabling caching can be expected to yield absurdly fast results. On a machine with 64GB of memory, you can see 60GB or so used for caching (if allowed)! You can easily see 3GB/sec read speeds with caching enabled (after having written it).

Even though DiskTester bypasses the Unified Buffer Cache, the system might still coalesce very small sequential writes, yielding results near the peak speed of the drive. Slightly larger transfers might then plummet in speed. Since this is “real world”, the results are in fact what you need to know, more relevant than the raw drive speed by itself. At any rate, this behavior cannot be disabled.

Drive caching

Hard drives vary in the amount of onboard cache. In general, any test that transfers an amount of data less than twice the size of the onboard cache is testing “burst speed”, which is relevant, but highly misleading for sustained workloads. A test size of 40X the drive cache size is recommended eg 1GB total I/O for a drive with a 32MB cache. Quadruple that for a 4-drive striped RAID.

Solid state drives have a variety of caching algorithms. Some use no cache, and some use a lot of it.

Random transfers

For random tests (the default), DiskTester uses unique random offsets; no two transfers are at the same location. The read phase uses the same locations as the write phase, in the same order by playing back the same series, but with reads instead of writes.

Test size

The test size spans a test file of this size:

(number of transfers) * (transfer size) + (gap size) * (number of threads -1)

DiskTester emits all relevant information as part of the test.

Key gotchas

Due to drive caching as well as optimizations in Mac OS X itself, very small write transfers (1K - 8K) can actually run much faster than larger ones (16K on up). Be careful not to assume too much based on one test!

Tips for testing with the iops command

The following will help get the most reliable results:

Command line usage

The command line allows greater control. Please use disktester help fill-volume for details on the available options. In general, there is little reason to use the command line.

iops
    [--xfer|-x <size[K|M|G|T]>]  "4M"
    [--transfers|-t <count>]     "100000"
    [--sequential|-s]            "true"
    [--threads|-n <count>]       "1"
    [--caching|-c]               "true"
    <volume-name>

Examples

Considerable variation is possible. In most cases, random testing is advised, since various other DiskTester command are designed for sequential I/O. Random is the default.

All examples below use volume “Scratch”; your volume is likely to have a different name. Use quotes if the volume name contains space(s), or eliminate the space(s).

Test random access speed with the default transfer size :

disktester iops Scratch

Test random speed with 1K transfer size and 1 million transfers:

disktester iops --xfer 1K --transfers 1000000 Scratch

Test sequential speed with 4K transfer size:

disktester iops --sequential --xfer 4K Scratch

Test speed with 2K transfer size and 4 simultaneous I/O threads:

disktester iops --xfer 2K --threads 4 Scratch

Test random access speed with the default transfer size (ultra fast reads):

disktester iops --caching Scratch

Previous page: run-sequential
Next page: run-streams