Swap Token - Advanced Operating Systems and Kernel Applications - page 104

Figure 12. The memory performance of two vortex programs (vortex1 and vortex2), each with a different
input, in their concurrent execution managed with swap token.

spike narrow after taking the token, while gzip
sharply reduced its RSS during this short period
of time. Vortex finished its execution before the
second RSS spike of gcc arrived. Process gcc
finished its execution without major page faults
after another 42 seconds. The execution times
of both processes are significantly reduced by
the swap token compared with the ones with the
original Linux LRU. The slowdown of gcc is 1.85
(a reduction of 67%), and is 1.54 for vortex (a
reduction of 54%). The page fault reductions for
gcc and vortex are 95% and 79%, respectively.
Figure 12 presents the memory performance
measured by MAD and RSS of two concurrent
running programs vortex1 and vortex2 when the
swap token is applied. At the execution time of
433 rd second, both processes started page faults
due to memory shortage. The token was taken by
vortex1 after this moment. Figure 11 shows that
the process then quickly built up its working set,
reflected by its climbing RSS curve after the token
was taken, while vortex2 continuously fluctuated
its RSS during this period of time. Vortex1 with
the token smoothly finished the execution and left
the system at the execution time of 668 th second.
Vortex2 then immediately obtained the needed

memory space, reflected by the sharp increase
of its RSS, and finished its execution without
major page faults after another 161 seconds. The
execution times of both programs are significantly
reduced by the swap token compared with the
ones in the original Linux LRU. The slowdown
of vortex1 is 1.95 (a reduction of 46%), and is
2.08 for vortex2 (a reduction of 38%). The page
fault reductions for vortex1 and vortex2 are 93%
and 63%, respectively.
Figure 13 presents the memory performance
measured by MAD and RSS of two concurrently
running programs LU1 and LU2 when the swap
token is introduced. In the first spikes of both
LU1 and LU2 processes after a few seconds of
executions, both processes started page faults due
to memory shortage.
The token was taken by LU1 after this moment.
Figure 12 shows that LU1 quickly built up its
working set, reflected by keeping its RSS curve
very similar to its RSS curve in the dedicated
environment after taking the token, while LU2
could only obtain a moderate amount of RSS
during this period of time. Process LU1 with the
token ran smoothly. In the last 25 seconds of the
execution of LU1 , its RSS curve was lowered