[HN Gopher] Tip: Time-Proportional Instruction Profiling [pdf]
___________________________________________________________________
Tip: Time-Proportional Instruction Profiling [pdf]
Author : gbrown_
Score : 13 points
Date : 2021-12-05 12:39 UTC (1 days ago)
(HTM) web link (folk.idi.ntnu.no)
(TXT) w3m dump (folk.idi.ntnu.no)
| cpu_architect wrote:
| I think there is a general problem with this line of research.
| This paper and others try to attribute execution time to specific
| instructions to provide feedback to software developers. But
| there's no way to verify that the resulting execution time
| breakdown is "correct." If the tool says 5% of execution time is
| spent in a particular ADD instruction, I can't just take out the
| instruction and verify this claim - the resulting program would
| be functionally different. Then how do I know if it's really 5%,
| 10%, or some other number?
|
| Not to say that there's no value in this research area; I would
| like the processor to provide feedback on where the performance
| bottlenecks are in my program. I would rather the researchers
| focus on generating meaningful optimization suggestions rather
| than just a table of numbers, though.
| efferifick wrote:
| Take a look at coz the causal profiler:
| https://github.com/plasma-umass/coz
|
| From what I know this is the closest project that fits your
| description.
___________________________________________________________________
(page generated 2021-12-06 23:02 UTC)