ScaleMM2 with MARS: +70% performance boost :-)

[Andrea Magni 75]

Just made a quick test using ScaleMM2 ( https://github.com/andremussche/scalemm ) with MARS (MinimalConsole demo, https://github.com/andrea-magni/MARS/tree/master/Demos/MinimalConsole) and it seems to push performances something like from 100 to 170% 🙂

If you are using MARS and want to give it a try, let me know how it works for you.

[David Heffernan 2345]

Doesn't this suggest that reducing the amount of heap allocation would lead to much bigger gains? 

[Andrea Magni 75]

1 hour ago, David Heffernan said:

Doesn't this suggest that reducing the amount of heap allocation would lead to much bigger gains? 

I am not actually an expert on this so low level topics but I am sure there is a big room for optimization in MARS.

So far I always focused on functionalities and ease of use and there are a couple of spots I know I can easily optimize. It's on my todo list but not really a priority at this very moment.

However, any help would be greatly appreciated, just in case somebody is willing to. 😉

[Stefan Glienke 2002]

It's easy enough to profile where the mass of heap allocations are coming from by using FastMM full debug and log the callstacks.

Or run under SamplingProfiler and check what hits the memory routines most.

Edited November 27, 2018 by Stefan Glienke

[David Heffernan 2345]

The point is that faster memory allocation treats the symptom rather than the cause. Such a large boost in performance implies that memory allocation is dominating execution time, at least in this benchmark. 

 

That in turn implies that reducing the amount of memory allocation could give far greater boosts.

[Andrea Magni 75]

@David Heffernan so I guess the fact I was just brute forcing 10k requests with ab.exe (Apache Benchmark) matters as the request served was a simple hello world thing. The setup time for the execution should be way more than execution time (physiologically).

 

It was just a simple test, but I am open to better benchmarking. 

 

Thanks

[Andrea Magni 75]

4 hours ago, Stefan Glienke said:

It's easy enough to profile where the mass of heap allocations are coming from by using FastMM full debug and log the callstacks.

Or run under SamplingProfiler and check what hits the memory routines most.

Thanks for the suggestions, I will give them a try but here we are a bit out of my comfort zone... 

 

Thanks @Stefan Glienke!