This framework can create a shadow workload (aka clone) given an x86_64 binary as input. The shadow workload demonstrates similar memory access pattern as the original workload while pruning all the non-memory instructions. The output of this framework is a C++ source code that can be compiled with GNU C or clang (not tested on icc).
Run make from the root directory. Successful build should create the obj-intel64/MAPProfiler.so file.
The tool itself is a target for Intel PIN, and should be invoked as follows:
$ ./pin/pin -t obj-intel64/MAPProfiler.so -func <function_to_be_profiled> -out <output_code_file> \
-top <top_filtering_percentage> -step <sample_size> -maxinst <maximum_inst_to_profile> \
-- <target_binary> <target_binary_arguments>
Where,
function_to_be_profiled: The user can limit the profiling to a specific function. If specified,
the shadow workload only mimics that particular function. Requires the
target_binary to be compiled with debug flag. (default: main)
output_code_file: Name of the generated shadow workload code file. It can be compiled with
a C compiler to get the shadow workload binary. (default: genCode.cpp)
top_filtering_percentage: During top rank filtering stage, this specified amount of memory accesses
will be kept at minimum. (default: 95)
sample_size: Sample size or batch size of the code fragments in number of memory
instructions. (default: -1, which means only a single code fragment)
maxinst: Maximum number of instructions to profile. This includes all instructions,
not just memory instructions. (default: -1, no limit)
target_binary: The original workload binary.
target_binary_arguments: Command line arguments for the target binary, if any.
$ cd examples/scatter_gather
$ make
$ cd ../../
$ ./pin/pin -t obj-intel64/MAPProfiler.so -out shadow.cpp -top 97 -step 1000000 -maxinst 5000000000 \
-- ./examples/scatter_gather/scatter_gather
This will clone the scatter_gather benchmark with a batch size of 1 million memory instructions, upto a maximum of 5 billion instructions. The generated clone have the file name shadow.cpp. It can be compiled as follows:
$ g++ genCode.cpp -o shadow -mavx
The -mavx flag is necessary since the clone can contain AVX load/store instructions.
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent