This archive provides plots of the schedulability and overhead experiments presented in Sections V, VI, and VII of the paper:
There are nine main folders corresponding to different experiments described in Sections V, VI, and VII, respectively.
The folder V-emstada-schedulability
contains six types of plots, arranged in subfolders.
part-only
: schedulability for various processor counts (m
) when using only partitioning heuristics. For each processor count, there are a number of plots, as identified by a tag at the end of the file name:
*_perc.pdf
— schedulability for various task counts, in the range from 0% to 100% relative utilization.*_top25_perc.pdf
— the same information, zoomed in to 75% to 100% relative utilization.*_top15_perc.pdf
— the same information, zoomed in to 85% to 100% relative utilization.*_top10_perc.pdf
— the same information, zoomed in to 90% to 100% relative utilization._fewer
omit curves for some values of n
to reduce clutter.semi-part
: The same types of plots as in (1) for basic semi-partitioning heuristics.
semi-part-paf
: The same types of plots as in (1) for semi-partitioning with the PAF meta-heuristic.
semi-part-rp
: The same types of plots as in (1) for semi-partitioning with the RP meta-heuristic.
semi-part-both
: The same types of plots as in (1) for semi-partitioning with the RP and PAF meta-heuristics.
strategy-comparison
: For each scenario (choice of m
and n
), a comparison of four general strategies: partitioning only, basic semi-partitioning, semi-partitioning with the PAF meta-heuristic, and semi-partitioning with both the RP and PAF meta-heuristics. For each scenario, there are a number of plots, corresponding again to different “zoom levels.”
The folder V-individual-comparison
contains plots showing all individual heuristics for a few selected configurations. For each scenario (choice of m
and n
), there are many plots, both representing different “zoom levels” and the following different groups of heuristics.
group=ALL
: Similar to type (7) above, but also showing a curve for semi-partitioning with only the RP meta-heuristic.group=PART
: Showing the performance of the considered partitioning heuristics. The “any” curve corresponds to the union of all curves (i.e., it claims a task set schedulable if any of the individual heuristics succeeded).group=SEMI
: Comparison of the individual semi-partitioning heuristics, again with an any
curve showing their combined performance.group=PAF
: Comparison of the considered instantiations of the PAF heuristic, including an any
curve.group=RP
: Comparison of the considered instantiations of the RP heuristic, including an any
curve.The folder V-unc-schedulability
shows data from an experiment the augments the main schedulability experiments by confirming the main trends with a different task-set generator. UNC-style generators do not allow individual control over the total utilization and number of tasks; hence there are two experiment types wherein either the task count or the total utilization are controlled (and the respective other parameter varies dependently).
The main parameter besides the number of processors m
is the util
parameter, which determines the distribution from which per-task utilizations are drawn. Precise definitions may be found on page 297 of the first author’s dissertation:
The folder VI-context-switch-rates
contains plots showing data from an experiment that compares the upper bounds on the maximum context-switch rate under the different scheduling approaches, using the Emberson, Stafford, and Davis task-set generator. For each considered combination of m
and n
, the folder contains three plots:
No tag: The comparison as shown in the paper (partitioning vs. semi-partitioning vs. semi-partitioning with meta-heuristics vs. QPS vs. RUN).
_pfair.pdf
: A comparison of RUN and QPS with the earlier Pfair approach, assuming different quantum sizes (100µs, 500µs, and 1000µs).
_ffd-vs-wfd.pdf
: A comparison of the different semi-partitioning approaches.
The folder VI-emstada-granularity
contains many plots corresponding to the experiment described in Section VI in which a variable minimum subtask size is enforced during semi-partitioning decisions. The results are grouped into respective folders by general strategy (partitioning only, basic semi-partitioning, etc.). For each scenario and each strategy, a number of different “zoom levels” are provided. The experiments depicted by the graphs in this folder used the Emberson, Stafford, and Davis generator.
The folder VI-unc-granularity
contains he plots corresponding to the confirmation experiments for the preceding experiments using UNC-style task-set generators. The files are structured and named analogously.
The folder VII.A-stock-plugin-comparison
shows graphs corresponding to the experiments described in Section VII.A of the paper. Each graphs shows certain overheads in terms of cycles. There are three types of graphs:
No tag: Shows the CDF of various overheads, sometimes using a log scale, and sometimes using a linear scale.
percentiles-*.pdf
: Shows (sort of) the inverse of the previous graphs: the observed percentile overheads, ranging from 0 to 100.
p90-*.pdf
: A version of the previous graphs, zoomed in to the range from the 90th to the 100th percentile.
See the artifact evaluation instructions and the LITMUSRT overhead tracing tutorial for more information concerning the individual overheads.
The same information as in the prior folder, for the experiment discussed in Section VII.B that compares the overheads of the proposed scheduling approach (SP-RES) with prior implementations of the optimal RUN and QPS algorithms
Since the experiments discussed in Sections VII.A and VII.B had to use different workloads due to implementation limitations in RUN and QPS, the plots in the folder VII.X-comparison-A-to-B
provide a comparison of the overheads incurred but the SP-RES implementation under the two workloads. The trends show that the workload used for the first experiment (Section VII.A) generate more overhead than the workload used for the second experiment (Section VII.B): the CDF for the first experiment is shifted somewhat to the right, relative to the CDF for the second experiment.