Push-Based Prefetching in Remote Memory Sharing System - Cloud, Grid and High Performance Computing: Emerging Applications - page 280

not hit the local cache of the file system. It is
interesting that the hit ratios of remote and local
memory decease rapidly with the increasing of
user nodes proportion, however hit ratios of the
local buffer cache and prefetching buffer are
slightly increasing. We can infer that the increas-
ing overhead in Figure 2 is mainly due to the
insufficient memory in the entire environment.
However, the performance of our prefetching
algorithm would not decease in this case.

prefetching. We can also see from Figure 5 that
the hit ratio of prefetching buffer decreases rapidly
with the decreasing prefetching buffer, whereas
that of the local cache and the prefetching buf-
fer does not change very much. This lies on the
fact that the free local memory is constant. The
decreasing prefetching buffer causes the increase
of the local cache, which reduces the effect of
prefetching buffer.

Simulation 3 . The effect of trace library size

Simulation 2 . The effect of prefetching buffer
size The effectiveness of the PSP algorithm is re-

lated to the amount of user patterns contained in
the trace library. We set a maximum trace library
size and the old traces should be discarded. By
default, the trace library contains at most 3000
traces. Indeed, it is near optimal value in Figure
6 and Figure 7. Obviously, when trace library
becomes small, the performance of PSP drops
rapidly. This is because there are not enough
training data to get the right prefetching list.
However, the larger size of the trace library can
also decrease the performance of the algorithm.
Both the overhead and hit ratio in the case of 8000
traces of the trace library are the worst in Figure 6
and Figure 7. This situation is true with not only
the transmission overhead of large trace library,

The prefetching buffer is one of key factors in
our scheme. It shares the free memory capacity
with the file system cache, and its maximum size
is restricted. In this set of simulations, we let the
upper bound of prefetching buffer to be from 1/2
to 1/128 of the free physical memory capacity.
The overhead and hit ratio with prefetching are
reported in Figure 4 and Figure 5, respectively.
Similarly, there exists a critical proportion range
for the prefetching buffer. When the proportion
of the prefetching buffer is within the range from
1/4 to 1/16, the performance of the algorithm
changes rapidly. The proportion of 1/3 is close
to the optimal, since it does not take much local
memory, and thus has the good performance of

Figure 3. Hit ratio with different user node pro-
portion

Figure 2. Reading overhead with different user
node proportion