Physical I/O Characterization: Pablo Physical I/O Tracing
Facility
The Pablo Physical I/O Tracing Facility is a program
that can be used to reveal the correlation between application I/O requests,
which are handled via the operating system, 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. An awareness of the distributions of trace
data quantifying disk I/O performance can enable programmers to choose optimal
static scheduling algorithms. For example, if trace data for a given application
indicates that 90% of all read operations executed experienced a read
duration time of less than 1000 microseconds, it is reasonable to assume
that efforts to optimize those I/O operations requiring lower read duration
times will benefit overall application performance more than efforts to optimize
operations calling for longer read duration times. Trace data also can be used
in real-time to dynamically adjust scheduling methods, optimizing application
performance on the fly.
In
contrast to other Pablo libraries, which are invoked at the user
application level,
the Pablo Physical I/O Tracing Facility is employed using a patch to the Linux
OS kernel. This patch is exactly the same as the standard Linux kernel except
it has been
augmented with instrumentation
software, attached before and after I/O handling routines, to capture and record internal
data. So, rather than using a standard version of Linux, users of the Pablo
Physical I/O Tracing Facility employ a patched version of Linux to run their source code and
thereby generate trace data profiling disk activity. As the
application executes, the performance data captured is recorded in a histogram. Then performance analysts can use
this information to profile and analyze the correlation between application I/O
requests and physical I/O operations in an effort to understand where the OS's
I/O subsystem efficiency could be improved.
An understanding of the patterns through which I/O requests are filled
clarifies the optimization strategies best suited to improving the performance
of such applications. Tracing the physical behavior of disk drives, as they are
being used to support distributed and parallel applications, yields data that
can be analyzed to reveal correlations between application I/O requests and the
physical activity undertaken by disk drives during I/O operations.