The tremendous rise in microprocessor technology has offered high speed processors and has increased the processor-memory speed gap dramatically. On the other hand, real-time embedded systems often have a hard deadline to complete their instructions. Consequently, the design of cache memory hierarchy is a critical issue in embedded systems. This paper describes a simulation-based performance evaluation of typical cache design issues in embedded systems such as using split caches for data and instruction versus unified cache for data and instruction, cache size and associativity and replacement policy. The evaluation is done using SimpleScalar simulation tools based on its Alpha version. We select some benchmarks for this study based on some previous researches about the clustering of SPEC CPU2000 benchmark suite. The contribution of this work is identifying important parameters for cache design in general-purpose embedded systems. Our results show that the Pseudo LRU techniques for cache replacement, such as MRU can approximate LRU with much lower complexity for a wide variety of cache sizes and degree of associativities.