Pete's Log: comp arch graph-o-rama
Entry #1018, (School)(posted when I was 22 years old.)
Ah, yes, the joys of computer architecture.
I think we've been sitting in this lab for five or six hours now, just collecting data. Lots of data. Tables upon tables of data. We've taken some of this data, and turned it into a variety of pretty graphs. We've got bar graphs, 3d bar graphs, logarithmic graphs, pie charts, and 3d pie charts. And to add flavor, we've stolen frames from some of the better phd comics out there that make reference to graphs and their goodness.
So my goal has been to find a case in which scoreboarding wins over pipelining. I've come close. For a SAXPY loop, scoreboarding performs almost as well as pipelining. But I've tried adding functional units, increasing latencies, and all sorts of other fun, and scoreboarding fails to outperform pipelining. I ran one test for which scoreboarding outperformed pipelining, but was unable to reproduce the results, since I forgot to note the exact conditions. I know it was something to do with high memory latency (small cache), but for the life of me I can't reproduce it. I really wouldn't care too much, since scoreboarding really just sucks, but Dr Uhran seems to have a fetish for the thing, so I want at least one case in which it performs better, just to satisfy his odd views of the world.
This project has actually been kinda fun. Implementing the scoreboard simulator was kinda challenging. Not hugely challenging, but enough so to be entertaining. Scoreboarding is an interesting concept. It just doesn't work too well in practice, and I already covered it in Dr Kogge's class last semester. But working with Katie on this project has been fun, we've got the coolest presentation ever. 10.5 hours remain until presentation time ...