1 16. Subject and Achievement of Past Research

2

3 I have been involved in research in the domain of computer architecture since my entrance into ENS Cachan, a higher education institution dedicated to training academics.

4

5 My first research efforts, for my bachelor's degree, were undertaken in the computer architecture research group of Prof. André Seznec at IRISA Rennes. There, I examined whether the measurements from on-chip temperature sensors included in modern CPUs could be used to determine the approximate power distribution on the chip. This turned out to be difficult, as even in an idealized model sensors were unable to reliably distinguish separate areas with different power densities.

6

7 Following this, I worked on parallelizing the temperature simulator used for the previous research. This work was carried out in the High Performance Computing group at Zuse Institut Berlin. I implemented and compared parallel versions in several programming models. From this I took an appreciation of the advantages, inconvenients and trade-offs in development of the different models, in particular OpenMP, CUDA, and threads.

8

9 Having explored software options for improving performance, I chose a Master's thesis project that would give me a complementary perspective. In the reconfigurable computing group at IRISA Rennes, I developed a high-level synthesis tool that, given the description of an instruction set, can generate an optimized datapath for implementation as an application-specific instruction-set processor softcore in an FPGA.

10

11 I continued in this direction in my subsequent internship at Potsdam University, where I implemented an LLVM compiler analysis pass that finds frequent computation patterns suitable for implementation as an application-specific instruction in an extensible or reconfigurable processor.

12

13 In my PhD project in the embedded systems architecture group at TU Berlin under the direction of Prof. Ben Juurlink, I pursue my investigation of the frontier between hardware and software. In my first year I examined the VMS runtime framework and its theoretical foundations. I formalized the underlying principles and instrumented the VMS code to analyze its performance. This work is presented in a paper currently under review at the EuroPar 2013 conference.

14 Since then, I have implemented runtimes for several parallel programming models using the VMS framework, notably Cilk and OmpSs. Using the previously developed instrumentation, I analyzed the performance commonalities and different requirements of the various models, with a view to developing a set of hardware accelerators that can support many different runtimes.