It bothered me throughout lecture: how are you measuring "busy-ness"? I would imagine this is done by counting the average number of operations per time unit. If this is the case, then it is unclear what the higher "busy" bars on the right represent. Are they due to:
1. The threads having to do more work due to the new algorithm in place?
2. The threads BEING ABLE to do more work as a result of not having to wait as long to get the data they need to process?
3. Cthulhu?
kayvonf
@lament: I don't know the details of how this particular experiment was run, so I'll answer the question in a modern context. Most modern hardware has mechanisms for profiling the performance of the processor, often called performance counters. These counters can provide detailed measurements of processor utilization.
We've never done a lecture on use of modern performance monitoring tools, but we should. I hope to have the TAs do some research on this and put together a recitation about modern tools in the semester half of the semester.
Olorin
It wasn't totally clear to me what the static assignment was here. The cost-zones assignment was referred to as "semi-static", if I'm understanding correctly, but I'm not sure what the static one is. Can someone clarify? Perhaps I missed something?
kayvonf
The static assignment was upfront random assignment of bodies (particles) to processors.
Olorin
Ahhh, okay. Thanks!
kayvonf
@Olorin. Random static assignment does not perform better. The Y-axis of the graph is execution time, and so the left graph indicates lower performance.
Olorin
Oops. Misread the slide. Guess I should go get some sleep before commenting more.
uncreative
@kayvonf
A recitation on profiling tools would be really cool. Especially in CUDA, I found profiling my programs fairly difficult, and often wondered if there was a better way to do it than just timing with the Cuda timers.
kayvonf
@uncreative. I agree. We are going to do one this year, but unfortunately it won't be until after Spring Break. The TAs are just too busy helping with assignments right now. But we'll do one in a time frame that might be helpful for projects.
It bothered me throughout lecture: how are you measuring "busy-ness"? I would imagine this is done by counting the average number of operations per time unit. If this is the case, then it is unclear what the higher "busy" bars on the right represent. Are they due to: 1. The threads having to do more work due to the new algorithm in place? 2. The threads BEING ABLE to do more work as a result of not having to wait as long to get the data they need to process? 3. Cthulhu?
@lament: I don't know the details of how this particular experiment was run, so I'll answer the question in a modern context. Most modern hardware has mechanisms for profiling the performance of the processor, often called performance counters. These counters can provide detailed measurements of processor utilization.
We've never done a lecture on use of modern performance monitoring tools, but we should. I hope to have the TAs do some research on this and put together a recitation about modern tools in the semester half of the semester.
It wasn't totally clear to me what the static assignment was here. The cost-zones assignment was referred to as "semi-static", if I'm understanding correctly, but I'm not sure what the static one is. Can someone clarify? Perhaps I missed something?
The static assignment was upfront random assignment of bodies (particles) to processors.
Ahhh, okay. Thanks!
@Olorin. Random static assignment does not perform better. The Y-axis of the graph is execution time, and so the left graph indicates lower performance.
Oops. Misread the slide. Guess I should go get some sleep before commenting more.
@kayvonf
A recitation on profiling tools would be really cool. Especially in CUDA, I found profiling my programs fairly difficult, and often wondered if there was a better way to do it than just timing with the Cuda timers.
@uncreative. I agree. We are going to do one this year, but unfortunately it won't be until after Spring Break. The TAs are just too busy helping with assignments right now. But we'll do one in a time frame that might be helpful for projects.