"... In PAGE 6: ...Executing all 1100 processes on a single 1 GHz node took about 520 hours to complete. The resources shown in Table 3 produced the results in Table4 . The theoretical best time are calculated from dividing the time to execute on a 1GHz processor by the total computational power, as measured in total GHz.... In PAGE 6: ... We are able to do this because the application actually scales well with the speed of the processor. From the results in Table4 , we were able to achieve close to the approximated best time. Once again, this was within the time of the average process duration on the slowest node (56 minutes).... ..."

"... In PAGE 4: ... Characteristics of EHW applied to real-world applications. Table1 summarizes the characteristics of the published work on EHW applied to real-world applications. The applications are mainly in the areas of classifica- tion, control and parameter tuning.... In PAGE 6: ... More details about FPGAs and FPAAs applied as EHW are given in [19]. Lack of commercial hardware that can be applied as EHW may explain why as much as half of the works presented in Table1 is based on custom hardware. 3.... ..."

"... In PAGE 4: ... Characteristics of EHW applied to real-world applications. Table1 summarizes the characteristics of the published work on EHW applied to real-world applications. The applications are mainly in the areas of classifica- tion, control and parameter tuning.... In PAGE 6: ... More details about FPGAs and FPAAs applied as EHW are given in [19]. Lack of commercial hardware that can be applied as EHW may explain why as much as half of the works presented in Table1 is based on custom hardware. 3.... ..."

"... In PAGE 6: ... Evolution is undertaken while the circuit is in operation and this makes the circuit online adaptable. Table1 summarizes the characteristics of the published work on EHW applied to real- world applications. The applications are mainly in the areas of classi cation and control when complete circuit design is applied.... ..."

"... In PAGE 6: ... As it was the case for the micro-benchmarks, re-execution of I/O system calls during the replay phase will also here result in less trace data to be captured to allow for a deterministic replay. This is especially the case for the ls directory listing appli- cation (see the last rows of Table3 ) and also shows up as a general lower slowdown on other applica- tions. 5 Related Work Until now, input replay tools have been primar- ily used for testing graphical user interfaces (GUI).... ..."

"... In PAGE 13: ...Table4 are in seconds: for compile amp; run and control ow, we give the sum of the total compilation time and the total execution time; for lazy control ow compilation, no distinction can be made, and so the total time for compilation and execution is given. First, we compare control ow compilation with compile amp; run.... ..."

"... In PAGE 14: ...or Table 2). The experiments are performed on the ILP datasets from Table 2. Additionally, the query pack execution mechanism allows us to do experiments on larger datasets, such as the HIV dataset (DTP). The timings in Table4 are in seconds: for compile amp; run and control flow, we give the sum of the total compilation time and the total execution time; for lazy control flow compilation, no distinction can be made, and so the total time for compilation and execution is given. Additionally, we give for each dataset the share of query goals that are never reached by the query execution.... In PAGE 15: ... The impact of leaving out the copying step is illustrated in Table 5. These are the results of running Tilde on the same datasets as in Table4 . Without copying the goals, the execution time of control flow compiled code becomes slower than code executed using the classical approach.... ..."