Architecture Research Project (Project 2)

   November 2012
Su Mo Tu We Th Fr Sa
11 12 13 14 15 16 17 <- topic
18 19 20 21 22 23 24 <- working prototype (and Thanksgiving break)
25 26 27 28 29 30

   December 2012
Su Mo Tu We Th Fr Sa
 2  3  4  5  6  7  8 <- presentations
 9 10 11 12 13 14 15 <- final draft (and final exam!)

• Topic: In class on November 13, be ready to very briefly (a few sentences) describe your chosen project topic.
• Prototype: Turn in working code by midnight Tuesday, December 20.  It doesn't need a performance analysis or writeup by this point.
• Presentations: Prepare about fifteen minutes of clear, easy to understand material introducing us to the topic, showing what you did, and discussing performance.  441 students will present Tuesday, December 4; 641 students will present Thursday, December 6.  No lecture notes or slides are required, but clear evidence of preparation is!
  
• Final: Turn in your tuned code and cleanly summarized performance analysis by midnight Thursday, December 13.  641 students should turn in a *brief* (about 4 pages) technical paper summarizing selected prior work, their design, and the results of a performance analysis.
  

Possible Project 2 Topics

• Build an interesting circuit: extend your HW1 CPU, build a superscalar dependency detection unit, etc.
• Hardware performance analysis: benchmark some test programs that demonstrates some aspect of modern hardware, such as:
  
• Out-of-order execution (e.g., reorder instructions manually, compare to automatic reordering)
  
• Branch prediction and execution speculation (e.g., reverse-engineer x86 branch hardware, like compare always-taken branch performance with even-odd branch performance)
• Dependency tracking (e.g., benchmark performance benefit from decreasing dependency tree depth)
  
• Cache prefetching and out-of-order loads and stores (e.g., compare cached loads with cached loads matching a previous store)
• Define a new instruction set, with a software or circuit simulator.
• Write and benchmark some code to perform any interesting task quickly on a particular architecture:
• Use bitwise operations to do something simple faster, or do something simple in a fiendishly complex way.
• Use assembly language or your knowledge of branch prediction, caching, etc to improve the performance of some program.
• Write a dynamic binary translator for any architecture.
• Use SSE or AVX instructions to speed up some code with the power of SIMD.
• Use OpenMP or pthreads to speed up some code with the power of multicore.  (But you must get the right answer!)
  
• Using MPI or sockets to speed up code with the power of clustering.
• Use CUDA or OpenCL to speed up some code with the power of the GPU.