sio_ntap: Simulated I/O (sio) Manual Page - Network Appliance

Overview

Manual page for the Simulated I/O tool (sio_ntap).

What is sio_ntap?
What are the command line options?
Example Inputs
Sample Output
sio_ntap compared to the Unix dd command

What is sio_ntap?

sio (acronym for Simulated I/O) is a general purpose I/O load generator. It performs synchronous I/O's to the specified file(s). The main purpose is to generate various I/O loads while collecting some basic statistic. In general, sio allows the user to control:

Read / Write mix
Random or Sequential I/O patterns
Access in various block sizes
Access over a variable amount of file space (starting at offset 0)
Adjustable runtime (in seconds)
Single or multiple concurrent threads performing I/O
Access to one or more files or devices (e.g. raw devices)

After completing the specified workload, sio generates several basic statistics:

I/O's Completed per second
KB/s Transferred
Total I/O's completed over the measured interval

sio is meant to enable I/O performance testing without having to create large application structures (such as databases). For example, sio can "approximate" a TPC-C like workload by specifying (for instance) a 2 to 1 read/write ratio, of 4k transfer sizes, with the appropriate number of threads. While the emulation is not exact, the approximation provides valuable insight into I/O subsystem performance.

What are the command line options?

The command line options are expressed here (obtained by running sio with no command line options):

Usage: sio Read% Rand% Blk_Size File_Size Seconds Thread Filename [Filename]

where

Read %: Percentage of accesses that are reads (versus writes)
Rand %: Percentage of accesses that are random (vs sequential)
Blk_Size: Size of I/O requests that are issued
FileSize: Size of area to be accessed in the file(s) (can be <= to actual file size; same for all files)
Seconds: Run time (specified in seconds), minimum of 10 seconds (60 or more recommended)
Threads: Number of concurrent threads issuing I/O's
Filename(s): One or more files to target for I/O

optional options

-noflock: Turn off filelocks. On some host, filelocks prevent host caching.
-output -outfile: Send all output to "outfile"
-direct: Disable filesystem caching - available on solaris, linux and aix.

Input Examples

Example 1: 100% Random reads of 512 byte xfers to filename1, running for 60 seconds with 1 thread, accessing 1MB of the file
- sio 100 100 512 1m 60 1 filename1

Example 2: Half reads, half writes of random 4K I/O's, filename1, 10 seconds, 2 threads, 20MB of file accessed
- sio 50 100 4k 20m 10 2 filename1

Example 3: Sequential writes of 64K I/O's for 60 seconds against filename1 with 1 thread, 10MB of file accessed
- sio 0 0 64k 10m 60 1 filename1

Example 4: 100% Random reads of 512 byte xfers to filename1, filename2, filename3, running for 60 seconds with 32 threads, accessing 1GB of each of the files
- sio 100 100 512 1g 60 1 filename1 filename2 filename3

Sample Output

server% sio 100 100 512 1m 10 1 testfile
Version: 3.00

SIO_NTAP:
Inputs
Read %:		100
Random %:	100
Block Size:	512
File Size:	999936
Secs:		10
Threads:	1
File(s):	testfile
Outputs
IOPS:		1995
KB/s:		998
IOs:		19974

The first block of output is simply a restatement of the command line inputs. Under the "Outputs" column, there are three result variables.

IOPS (I/O's Per second): indicates average I/O's per second over the entire run.
TPUT (Measured in KB/s): indicates average KB/s transferred over the entire run.
IOS: Total I/O's completed over the length of the run.

Beware when making comparisons among platforms or technologies. The most common issue is comparing I/O's that are cached on the I/O device with some other measurement of unCached I/O's.

sio compared to the Unix dd command

Many times, "dd" is the first "performance test" that is run against a new I/O subsystem. As a water test, this is fine. As a real measure of overall performance, dd and sio differ in a number of ways:

Single threaded. dd is a single threaded test. While this provides some measure of throughput, it does not measure "maximum" throughput, since a single threaded requestor is heavily affected by I/O latency. sio allows both single thread and multiple thread workloads.
dd has a fairly restricitve data movement paradigm: read from one file write to another. sio allows simulation of a wide range of workloads, including sequential and random I/O's, read and write workloads, etc.
dd's run time is dictated by the file size. To obtain really long runs, dd needs either a large file or requires repetitive runs (making stats collection difficult). sio can run any amount of time on any size file. This simply makes testing sustained performance workloads easier.
dd's stats are somewhat limited. sio provides the three most useful stats in a concise form. Also, sio has the capability to collect more extensive statistics in certain environments.

Revised: July 01, 2004