This paper presents a version of skewed associative cache, called selective-way-access skewed associative (SSA), for embedded systems. The SSA illustrates less energy consumption in comparison with direct mapped and two-way/four-way set associative caches.

The work does not offer any new breakthrough ideas. Skewed associative caches have been proposed from the early 1990s. Seznec and Bodin, who the authors cited, talked about different skewed associative caches. The authors of this paper did provide different hashing functions for the skewed set (the second set for avoiding the cache conflict misses). Some new components, such as the lookup table to get to the sets, are also introduced. Since SimpleScalar tool sets have a cache simulator, I do not understand why the authors needed to dump the cache traces, and use Dinero to evaluate the cache behavior of the MediaBench benchmarks.

I did like the last table on page 889, Table 3. This table reports that the proposed cache organization (SSA) outperforms any other of the listed caches. What remains to be explained is the abrupt change in energy consumption between two-way and four-way set associative caches. I fully understand why four-way and two-way prediction caches perform better than their counterparts. Way-prediction caches reveal fewer cache misses, and therefore more accesses hit the cache, and do not need the second level cache access. Overall, having said that a great percentage of the energy consumed in embedded systems is spent in the memory hierarchy, it seems very rational to spend a lot of time on research to improve the energy consumed in the cache, or to improve the locality behavior of media applications, something that I think is worth exploring.

I found the English in this paper to be very poor. I had a hard time reading the text; I needed to read a paragraph several times to understand its point.