Swap Token - Advanced Operating Systems and Kernel Applications - page 95

vortex RSS curve after 760 th second in Figure
3. We believe the pages gzip lost are part of its
working set, because it had increased number
of page faults and tried to gain some allocation
back after then. Though vortex can take certain
memory spaces from gzip , it is unable to build
up its working set. This is because it also lost a
large number of false LRU pages when it tried to
build up its working set, which should not have
been lost considering the needs of vortex . Unfor-
tunately, we observed that the system ended up
with high page fault rates for both processes and
a low CPU utilization. We found that a process is
powerful to get additional memory allocation in
the global replacement policy when it has large
memory shortage between its RSS and its work-
ing set. However, when it gets more memory, it
becomes less powerful, and tends to lose memory.
For this reason we see the fluctuating RSS curves
for the concurrently running programs in the
system thrashing. Our experiments show that the
execution times of both programs are significantly
increased due to the page faults in the concurrent
execution. The slowdown of gzip is 5.23, and is
3.85 for vortex .
Figure 4 presents the memory usage behavior
measured by MAD and RSS of concurrently run-

ning programs bit-r and gcc . Gcc belongs to type
3 which has two spikes in MAD and RSS due to
its dynamic memory demands. For bit-r , its RSS
curve dropped sharply from 32,800 pages to about
16,500 pages at the 165 th second caused by the first
RSS spike of gcc at the same time. After the spike,
the RSS of gcc was decreased, which allowed bit-r
to regain its RSS. When the second RSS spike of
gcc arrived at the 365 th second, the RSS of bit-r
dropped again. However, this time the RSS of
gcc began to lose its pages at about 450 th second
before it could establish its working set.After that,
both programs exhibited fluctuating RSS curves.
The second spike requires only 7% more memory
demand than the first spike, which causes a much
longer execution delay. Consequently, program
gcc 's second spike of the MAD and RSS curves
were stretched to a delay of 357 seconds due to
page faults. During this period, there was a big
gap between the RSS and MAD, up to more than
20,000 pages. The experiments consistently show
that the execution times of both programs were
significantly increased due to the page faults in
the interaction. The slowdown of bit-r is 2.69,
and is 3.63 for gcc .
Figure 5 presents the memory behavior mea-
sured by MAD and RSS of concurrently running

Figure 5. The memory performance of gcc and vortex in their concurrent execution.