gather_scatter

Micro-benchmarking gather and scatter instructions

RQ.- Can we build an analytical model of the performance of gather and scatter instructions based on number (and set it maps to?) of cache lines touched under hot cache?

Results: conclusions

All raw data is under data directory.

Profiler

The profiler has been able to compile properly the code and generate a loop including only the instructions we want to measure, e.g.:

	call	polybench_papi_start_counter@PLT
	test	eax, eax
	jne	.
	mov $10000, %ecx
begin_loop:
	vmovdqa	ymm2, YMMWORD PTR .LC1[rip]
	vmovaps	ymm1, YMMWORD PTR [rsp]
	vgatherdps	ymm0, DWORD PTR [r14+ymm2*4], ymm1
	sub $1, %ecx
	jne begin_loop
    mov	edi, ebx
	vzeroupper
	inc	rbx
	call	polybench_papi_stop_counter@PLT

This way, the profiler is only including within the loop (begin_loop) those instructions we want to measure in the micro-kernel. The same example applies for vscatterdps. With these configurations, we generate 13122 variants or versions of the source code.

Analyzer

Our analyzer enables the exploration of the most promising features in the performance. The processing of data is, normally, an iterative process, where on each step we learn more about the system to characterize or model. Thus, in a first step, we explore data using a relational plot (see configuration file):

Gather performance seems pretty steady, regardless some outliers, but scatter performance seems to be more volatile, requiring further inspection.

Experimental configuration:

• Intel(R) Xeon(R) Silver 4216 CPU @ 2.10GHz
• Compiler: GCC 11.2.0
• MARTA Profiler Full Configuration file

How to reproduce:

NOTE: in this example we already provide the profiling data, under data directory, so the profiler step can be skipped. In case you want to run the profiling step, please be careful with paths and input files in the analyzer configuration file.

1. Profiling

   # place at the root of the project
   marta_profiler examples/gather_scatter/config_profiler.yml
   # or
   python -m marta.profiler examples/gather_scatter/config_profiler.yml
   

   This step should take around 2 hours (Intel Xeon Silver 4216). This will generate a CSV file with all measurements and dimensions of interest in examples/gather_scatter/src/marta_profiler_data. This will be the data used by the analyzer.

2. Analysis

   # place at the root of the project
   marta_analyzer examples/gather_scatter/config_analyzer.yml
   # or
   python -m marta.analyzer examples/gather_scatter/config_analyzer.yml
   

   This step is almost immediate (around 10 seconds), and will generate a series of plots and files with the result of the classification algorithms (decision tree) and the feature importance analyzer (random forest classifier).