Creating compression traces.

To record the actual effectiveness and time cost of compressing each page image, we created a set of compression traces. For each combination of compression algorithm and CPU, we created a trace recording how expensive and how effective compression is for each page image in the reduced page image trace. Since we have 6 test programs, 3 compression algorithms, and 3 CPU's, this resulted in 54 compression traces.

The tool that creates compression traces is linked with a compressor and decompressor, and consumes a (reduced) page image trace. For each trace record in the page image trace, it compresses and decompresses the page image and outputs a trace record. This record contains the page number, the times for compressing and decompressing the page's contents at that moment, and the resulting compressed size of the page. Each page image is compressed and decompressed several times, and the median times are reported. Timing is very precise, using the Solaris high-resolution timer (all of our compression timings were done under the Solaris operating system). To avoid favorable (hardware) caching effects, the caches are filled with unrelated data before each compression or uncompression. (This is conservative, in that burstiness of page faults will usually mean that some of the relevant memory is still cached in the second-level cache in a real system.)