Raw data for the Memory Exclusion Paper
New Data (8/12/96)
In a footnote to the paper submission, we say: `` There is an
apparent anomaly in Figure 3 that should be explained. This is the
fact that in NNET, SOLVE and CELL,
incremental checkpointing improves the overhead even though
the checkpoint sizes are the same. This should not be the case ---
incremental checkpointing should actually penalize performance here
because since the same amount of data is being written to disk, and
there is extra overhead processing SEGV signals. This anomaly arises
because the operating system buffers user writes that do not fall on
page boundaries, and libckpt does not take account of this
fact in non-incremental checkpointing. This bug will be fixed
shortly, and new data will be taken so that this anomaly is no longer
We have fixed this bug so that all writes occur on page boundaries.
The new figure 3 is here,
and the new raw data files are below.
Note that there are more runs per datum as well in these tests.