Skip to content

yhuai/ioscm

 
 

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

58 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

#Introduction# I/O Stream Congestion Meter is a benchmark to measure the effect of multiple concurrent I/O streams on one storage system. Each I/O stream in the benchmark is a squence of interleaving I/O operations and fake computation (Idle). It is part of an effort to understand the performance degradation (throughput, skew, fairness, ...) caused by I/O congestion for file systems like HDFS, PVFS, and ext4. Its features include:

  • Computation-I/O pattern per I/O stream group
  • Fast synchronization for small I/O streams
  • Supports hundreds of concurrent read/write I/O streams

#An Example#


Figure 1 A Case of I/O Congestion

#Build# ant -f build.xml dist

#HowTo#

  • xml file in conf/ defines the type and parameters for each I/O stream (group) in the testing
  • tool/ioscm-conf.py is a tool to specify parameters in configuration files
  • scripts in tests/ give detailed usage cases

#TraceReplayer# A TraceReplayer takes an I/O trace file as input, and performs I/O accordingly. The format of an I/O trace file is

  • Trace line format: "offset(bytes, sector)|size(bytes)|operation|wait_after_operation(seconds)"
  • There're 4 types of operations: R(RandomAccessFile.read), r(blocking read), W(write), w(write and sync)

To launch multiple TraceReplayers at the same time, one can launch an I/O stream group called TraceReplayerBatch with the following configuration:

	<Stream label="TraceReplayerBatch" type="TraceReplayerBatch">
    	<period>4</period>
    	<dataDir>/expr/data</dataDir>
   		<traceDir>/expr/trace/WebSearch1.spc_s</traceDir>
   		<isBlock>true</isBlock>
	</Stream>

In the above case, ioscm will launch an I/O stream for each trace file within /expr/trace/WebSearch1.spc_s/. It is required that each trace file has a data file with the same file name inside /expr/data/ before the test starts. In addition, all streams will halt after 4 seconds, and the offset unit is block(512Bytes). For detailed usage case, please refer to tests/TraceReplayerBatchTest.sh

#TraceReplayer7#

##Environment## Set the home of JAVA 1.7 in reference to conf/ioscm-env.sh

##Change from TraceReplayer## TraceReplayer7 requires Java1.7 and offers a slightly different set of I/O operations

  • R(asynchronous read), r(blocking read), W(asynchronous write), w(blocking write and sync)

##Configuration## More options are included in its configuration file

	<Stream label="TraceReplayer7Batch" type="TraceReplayer7Batch">
    	<period>4</period>
    	<dataLocation>/dev/raw/raw1</dataLocation>
    	<dataIsDir>false</dataIsDir>
    	<traceDir>/home/siv/expr/trace/tiny/</traceDir>
    	<blockSize>1</blockSize>
    	<intervalScaleFactor>1</intervalScaleFactor>
    	<AIOPoolSize>128</AIOPoolSize>
    	<SyncMode>SYNC_ALL</SyncMode>
	</Stream>

Here <traceDir> gives the path of the folder that contains all the I/O trace files. For each trace file, the replayer will then launch a thread for replaying. Besides, <dataLocation> specifies where to access the data, and <dataIsDir> indicates whether the specific locaiton contains one singel file, or a set of files. If false, all launched streams will access the same file concurrently; otherwise, for each trace found in <traceDir>, there must be one data file using the same file name within <dataLocation>.

In addition, two options are supplied to let the user slightly twist the I/O trace at runtime. <intervalScaleFactor> allow user to scale the wait_after_operation(seconds) in trace as needed. For instance, assigning a value 2 would simply double the waiting time. <SyncMode> is another handy option. Value SYNC_ALL change all opeations to synchrounous; and value ASYNC_ALL asynchrounous.

There is also one AsynchronousFileChannel specific option, <AIOPoolSize>. The value of this tag specifies the size of the thread pool for asynchronous I/O.

##Raw Device Support## Raw Device access is a way to bypass OS cache buffers for Java programs. To enable such functionality, the jar package needs to be launched with -Dsun.nio.PageAlignDirectMemory=true argument. More details can be found in script tests/TraceReplayer7Batch.sh.

#Combine Different Stream Groups# To corun different groups of streams, one only needs to put their group configurations together. conf/UnlimitedWriterReaderBatch.xml gives an example:

<Stream type="UnlimitedWriterBatch" label='UnlimitedWriterBatch'>
    	<number>4</number>
    	<interval>0</interval>
    	<period>16</period>
    	<rsize>65536</rsize>
    	<path>/tmp/</path>
</Stream>
<Stream type="UnlimitedReaderBatch" label='UnlimitedReaderBatch'>
    	<number>4</number>
    	<interval>0</interval>
    	<period>16</period>
    	<rsize>65536</rsize>
    	<path>/tmp/</path>
</Stream>

Note that labels on different groups should be different.

#Log# By default, ioscm dump logs to terminal at runtime. There are two types of messages.

  • Message on progress. It indicates the state of current I/O stream. It nomarmlly comes with a state message followed by the stream ID. Examples inlcude:

    Sync.start ID

    Sync.end ID

  • Message for profiling. It has the format:

    MeterTimer tag label ID start_time end_time duration val1 val2 [op]

#People#

About

I/O Stream Congestion Meter

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Java 54.8%
  • Shell 35.2%
  • Python 8.7%
  • MATLAB 1.3%