Using the Pablo Physical I/O Facility: A Tutorial
The Physical I/O Facility is a software toolkit that, when
coupled with application I/O instrumentation, helps reveal the correlation
between application I/O requests and physical I/O operations. This is
significant because physical I/O patterns are strongly affected by data striping
mechanisms, file system policies, and disk hardware attributes. Understanding
how the operating system translates application I/O requests into physical disk
operations can aid in optimizing file policies and data distributions for higher
performance.
The Physical I/O Facility captures I/O data quantifying the
performance of SCSI disk drives supporting applications run on a Linux
system. Unlike other Pablo I/O characterization tools, it is not a
user-level library that can be substituted for standard language or operating
system I/O calls. Rather, it consists of a patch to the Linux kernel and a set
of user-level tools. The Linux kernel patch performs exactly the same function
as its standard Linux kernel counterpart, but it has been augmented by instrumentation
software, attached before and after reads and writes, to capture internal
data. Programmers run their applications in a Linux environment that has been
modified with the Physical I/O Facility's Linux kernel patch. During program
execution, performance data, captured by the instrumentation, is recorded in a
set of histograms. Analysts study these histograms to understand how the
Linux device drivers translate application I/O requests into physical disk
operations. This tutorial includes: instructions
for patching the Linux kernel and instructions
for compiling and using the user-level tools, as well as access to an end-to-end
example of performance analysis.
All physical input/output requests pass through an operating system's
device drivers. These drivers define the interface between file system
services and I/O devices, isolating the idiosyncrasies of specific
peripherals behind standard interfaces. By
instrumenting device drivers, analysts can reveal the activity patterns of
all I/O requests generated by a file system.
Quantifying the behavior
observable at the SCSI device drivers, the Physical I/O Facility
tracks the execution of reads and writes, capturing the quantity, size, and
duration of these operations and the length of time spent waiting, in the
device driver queue, to access the disk.
|
Data captured on
Read operations
|
Data captured on
Write operations
|
| Quantity (# bytes) |
Quantity (# bytes) |
| Duration in Device
Queue |
Duration in Device
Queue |
| Duration of Entire
Read Operation |
Duration of Entire
Write Operation |
Pablo Physical I/O Facility users apply a kernel patch to supplement the
standard Linux kernel with the instrumentation required to generate physical
I/O performance metrics. The various patches included in the Pablo Physical
I/O Facility are based on official versions of the Linux kernel source and
patch releases. The Linux kernel
distributions can be downloaded from http://www.linux.org
To apply the kernel patch, users first choose the patch file appropriate
for the kernel version being used. At the top directory of the Linux kernel source tree, execute
the following command:
| Command |
% patch
-l -p0 <patch filename with the full path> |
| Example |
% patch -l
-p0 </tmp/phIO-patch-2.2.16 |
After this, compile and install the kernel.
Finally, reboot the machine with the new kernel.
For instructions of how to build the Linux kernel, refer to Linux
documentation at http://www.linuxdoc.org/HOWTO/Kernel-HOWTO.html
The
Physical I/O Facility GUI Toolkit can be used to initialize or reset the
physical I/O data structures in the kernel, and to retrieve physical I/O data
from the kernel in a one-time fashion or periodically. It can also plot physical
I/O data statically, from the data file, and dynamically, as it retrieves the
data in real time. The
user-level GUI is programmed in Java and it can be run on any systems that
support the Java Virtual Machine. The GUI was developed using Sun JDK version
1.3. However, our tests indicate that it runs smoothly in Sun JDK version 1.2.2
environment as well.
The
top menu bar has three buttons:
-
File
button for opening existing physical I/O data file, saving physical I/O data
to a file, saving histograms to JPEG files, and exiting GUI.
-
Tools
button for launching physical I/O user-level tools.
-
Help
button for help on using the GUI.
-
Open: Opens the specified physical I/O data
file.
-
Save Data: Saves the retrieved physical I/O
data (either once or periodically) to a file with the previously specified
name.
-
Save Data As: Similar to "Save Data"
button, but requires that the user designate the file name.
-
Export JPEG: Saves the currently plotted
figure to a jpeg file with the previously specified name.
-
Export JPEG As: Similar to the "Export
JPEG" button, but requires that the user designate the file name.
-
Print Image: Print the current figure to a
printer specified by user.
-
Exit: Exit
the GUI.
-
Initial
Data: Initializes
and resets all Physical I/O data structures in the kernel. This function
requires the user to input the SCSI driver name, the bin sizes for the read
request size (in bytes), the read driver queue time (in
microseconds), the read request duration (in micro-seconds), the write
request size (in bytes), the write driver queue time (in
micro-seconds), and the write request duration (in micro-seconds).
-
Get
Data: Gets
the current Physical I/O trace data for the SCSI driver specified by the
user. When the command is launched, the data retrieved will be
automatically plotted in the plotting area.
-
Get
Data by Interval time: Gets
sequential Physical I/O trace data from the SCSI driver specified by the
user. The user must also designate the interval time and the number of
iterations to capture. When the command is launched, the data retrieved in
each time interval will automatically be plotted sequentially in the
plotting area.
-
About:
Shows the
version and overview description of PIOF.
-
Help
Topics: Provides an index of topic covered in the PIOF user
help library.
The
Plotting Area has three parts.
On
the top, there are 6 buttons denoting 6 metrics:
-
(r)
request:
for Read request size
-
(r)
queue: for Read request
driver queue time
-
(r)
duration:
for
Read request duration
-
(w)
request:
for Write request size
-
(w)
queue: for Write request
driver queue time
-
(w)
duration:
for
Write request duration
Clicking
one of these buttons will plot the distribution of trace data characterizing
that parameter in the drawing area.
In
the middle of the Plotting Area, physical I/O data will be drawn for the
specific metric chosen by the user. The X-axis represents the metrics, for
example Read
request size.
The Y-axis represents the number of occurrences in each histogram bin of the
specified metric.
At
the bottom, there are two buttons: Histogram and Line. Users can
choose the drawing format by clicking on either of these buttons. The Histogram
button is for representing the data in traditional histogram fashion using bars. The Line button
draws the data in a curve
.
When an disk operation occurs, the Physical I/O Facility generates an
entry in a histogram, a table.
Before
running the GUI, set the environment variable “CLASSPATH” to the path
housing the class files of executable Java files. For example.
% setenv CLASSPATH /usr/local/java/classes
To
run the GUI, type “phyio”
at the command line in the directory in which the tool is installed.
% phyio
Download
the Pablo Physical I/O Facility User Guide