Implementation and Architecture Effects

In the past sections we only showed results for our SPARC 168 MHz machine. As expected, the faster SPARC 300 MHz machine has a lower compression and decompression overhead and, thus, should perform better overall. The Pentium Pro 180 MHz machine is usually slower than both SPARC machines in compressing and uncompressing pages (not unexpectedly as it is an older architecture--see also out later remarks on memory bandwidth).

  

Figure 5: A SPARC 168 MHz usually has better performance than a Pentium Pro 180 MHz, while a SPARC 300 MHz is significantly better than both. Nevertheless, the Pentium Pro 180 MHz is much faster for a hand-optimized version of the LZO algorithm, sometimes surpassing the SPARC 168 MHz.

Figure 5 shows three of our test programs simulated under WKdm and LZO in all three architectures. For WKdm, the performance displayed agrees with our observations on machine speeds. Nevertheless, the performance of LZO is significantly better on the Pentium Pro 180 MHz machine than one would expect based on the machine speed alone. The reason is that, as pointed out earlier, the implementation of LZO we used on the Pentium Pro is hand optimized for speed in Intel x86 assembly language. Perhaps surprisingly, the effect of the optimization is quite significant, as can be seen. For ghostscript, for instance, the Pentium Pro is faster than the SPARC 168 MHz using LZO.