Latency heat maps
De TAIl oF FIGuRe 7 FRoM PAGe 53
Given the ability to trace latency at arbitrary points of interest, the problem
becomes effective visual presentation
of this data. Busy systems can be processing hundreds of thousands of I/O
events per second, each one providing
a completion time and I/O latency. One
approach is to summarize the data as
average and maximum latencies per
second, which can be presented as line
graphs. While this would allow average
latency to be examined over time, the
actual makeup or distribution of that
latency cannot be identified beyond a
maximum, if provided.
To examine a distribution over time,
visualizations such as heat maps may
be used. The use of heat maps in sys-
tem observability tools has been infre-
quent, with some appearances to map
the access pattern of disk I/O. An exam-
ple of this is taztool (1995), which dis-
plays a heat map showing time on the
x-axis and disk I/O offset on the y-axis,
allowing random and sequential disk
I/O patterns to be identified by visual-
izing the location of disk I/O. 5
For the latency heat map to be most
effective, the time and latency ranges
represented by each pixel should be
sufficiently large to allow multiple I/O
operations to fall within them. This allows darker shades to be selected and
patterns shown by different shades to
be observed. If the ranges are too small,
many of the pixels may represent only
one I/O, and much of the heat map
may appear in the same color shade; it
may also reduce the likelihood that adjacent pixels are shaded, and the heat
map may look more like a scatter plot.
The range of possible color shades
from light to dark may be applied to
each heat map generated. This can be
applied linearly: the pixel with the most
I/O is assigned the darkest color, and
all other pixels are given a shade that
is scaled from the darkest I/O count.
A drawback with this approach is that
