Alleviating the Thrashing by Adding Medium-Term Scheduler - Advanced Operating Systems and Kernel Applications - page 128

Figure 3. a. SPEC's Number of Swaps; b. SPEC's Execution Time

the swap area was necessary, because just 14MB
out of each Matlab process was physically in the
internal memory and the other memory allocations
(except of the shared allocation) of the Matlab
processes were in the swap area. We preferred no
to reshow the results that are almost the same as
Figure 2a and 2b and instead to show in the next
figures different benchmarks results.
Figure 3a and Figure 3b show the performance
of the medium-term scheduler vs. the Linux kernel
using the tests of SPEC cpu2000 benchmarks. The
prefix 3 (or 2) before the test name indicates that
we iterated the test 3 (or 2) times. Sometimes we
divided the numbers by some constants in order
to fit the data to the scale of the diagram. These

constants are denoted as Test/Constant. When
we used more than one test, we added a '+' sign
between the names of the tests.
When each group contains just a few memory-
consuming processes, the idle task might be
invoked too often, even though there are other
processes in other groups that can be executed.
This can reduce the time saved by eliminating the
thrashing effect. When a test has large memory
allocations and is executed in a different group,
the results will not as good as when executing
several smaller SPEC tests concurrently in one
group. A higher idle time will be emerged when
the content of each group is just one process; thus
the results of Figures 3a and 3b are not as good