Re: [Valgrind-developers] How can I improve the performance of my memcheck customization?

[Paul F. <pj...@wa...>]

On 18-01-25 12:09, Enzo Bonansea wrote:
> Hello,
> 
> For my computer science thesis, I need to instrument programs in order 
> to know which floating point values are being stored to buffers whose 
> size is greater than a certain threshold. For doing that, I modified the 
> tool memcheck in this public repository <https://github.com/ 
> enzobonansea/floating-point-vectors-inspector>; you can find my 
> modifications by searching for "memlog" in the code.
> 
> The customization works as I expect, giving me the floating point values 
> that I need. Now, I want to optimize my code since it's CPU intensive 
> and makes the instrumented applications too slow. For instance, I ran 
> the memcheck customization against the SPEC2017 tool called lbm which 
> took originally a few seconds to run, but under the tool it takes more 
> than 2 minutes with the CPU at 100% all the time.
> 
> If you see the code, the cause of the performance degradation is 
> evident: for every store I'm calling the dirty helper log_store, which 
> means that I'm calling that C function millions of times.
> 
> I used Valgrind's memcheck several times as a user, but 2 months ago was 
> my first time seeing the tool's source code. So, since I'm new on this 
> and you guys have a lot of experience working on the project, maybe you 
> find my problem trivial and you can have a good idea of how I can 
> improve the performance of my customization of memcheck.


Hi

Do you really need to base your features on memcheck?

I imagine that you would get better performance if you wrote a new tool 
(perhaps starting with lackey).

After that, use a cache and write to file as little as possible (but I 
from what I see you do that already).

A+
Paul