/HyFiSS

HyFiSS: A Hybrid Fidelity Stall-Aware Simulator for GPGPUs

Primary LanguageC++

DOI

The source code for the simulator presented in our paper at the "MICRO 2024" conference.

Paper Details

  • Title: HyFiSS: A Hybrid Fidelity Stall-Aware Simulator for GPGPUs
  • Conference: MICRO 2024

Prerequisites

  • C++ Compiler (g++), the version we use is g++ (GCC) 7.3.1 20180303 (Red Hat 7.3.1-5). All compilers that support the C++11 standard will do well. You can install by this way: 
    yum install g++
  • Make, the version we use is GNU Make 4.2.1. You can install by this way: 
    yum install make
  • Python, the version we use is Python 3.7.0. We recommend you to use Anaconda to manage Python packages, and the guide to install Anaconda is available here.
  • MPICH, the version of MPICH we use is v3.3.2 and it is available here. Download the source code, uncompress the folder, and change into the MPICH directory. 
    wget https://www.mpich.org/static/downloads/3.3.2/mpich-3.3.2.tar.gz
tar -xzf mpich-3.3.2.tar.gz
cd mpich-3.3.2
    After doing this, you should be able to configure your installation by performing ./configure. 
    ./configure
    If you need to install MPICH to a local directory (for example, if you don't have root access to your machine), type ./configure --prefix=/path/to/your/path. When the configuration is done, build and install MPICH with make && sudo make install. After installing MPICH, you also need to set the environment variable: 
    export MPI_HOME=/path/to/mpich-3.3.2
export PATH=$PATH:$MPI_HOME/bin
export MANPATH=$MANPATH:$MPI_HOME/man
export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:$MPI_HOME/lib
    After this, you should be able to type mpiexec --version or mpirun --version and see the version information of MPICH.
  • Boost Library, the version we use is v1.83.0, and it is available here. Download the source code, uncompress the folder, and change into the Boost directory. 
    wget https://boostorg.jfrog.io/artifactory/main/release/1.83.0/source/boost_1_83_0.tar.gz
tar -xzf boost_1_83_0.tar.gz
cd boost_1_83_0
    After doing this, you should be able to compile the Boost library. 
    ./bootstrap.sh -prefix=/path/to/boost
    Where /path/to/boost is the path you want to install Boost in. As explained in the boost installation instructions, running the ./bootstrap.sh from the boost root directory will produce a project-config.jam file. You need to edit that file and add the following line: 
    using mpi ;
    When the above is done, install Boost with sudo ./b2 install， and this will allow you to install the Boost library in the /path/to/boost directory. After installing, you also need to set the environment variable: 
    export BOOST_HOME=/path/to/boost
export CPLUS_INCLUDE_PATH=$BOOST_HOME/include:$CPLUS_INCLUDE_PATH
export C_INCLUDE_PATH=$BOOST_HOME/include:$C_INCLUDE_PATH
export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:$BOOST_HOME/lib

Building the Simulator

  1. Clone the repository:

    git clone https://github.com/RepositoryAnonymous/gpu-simulator.git gpu-simulator

  2. Change to the project directory:

    cd gpu-simulator

  3. Build the simulator:

    make all -j

The compiled simulator executable is called gpu-simulator.x, and is just in the current directory.

Clone the Benchmark Suit

The benchmark application suit we used for validation consisted of 43 applications with a total of 1784 kernels. We evaluate them on a real NVIDIA QUADRO GV100. These applications are from the cuBLAS library, the heterogeneous computing benchmark suite PolyBench and Rodinia, the fluid dynamics benchmark LULESH, the deep learning basic operator benchmark suite DeepBench, the DNN benchmark suite Tango, the unstructured mesh mini-app benchmark PENNANT. For all selected applications, we select their first 100 kernels for evaluation. All workloads are compiled using CUDA 11.8 with the compute capability sm70 for the Volta architecture. You can clone this benchmark suit we used from here.

  1. Change to the simulator project directory:

    cd gpu-simulator

  2. Clone the benchmark suit repository:

    git clone https://github.com/RepositoryAnonymous/simulator-apps.git apps

For guidance on compiling these applications and getting their traces on real hardware, please refer to here. And this repo also provides the guide to obtain experimental results and reproduction images of our paper. All compilation scripts are being sorted out and will be released in a few days (current time: August 1, 2024). We have evaluated more than 50 applications from many different benchmark suites, so writing a compile script for them all will take some time.

Preprocessing the Traces

Please refer to our reply here.

Running the Simulator

To run the GPU simulator, use the following command:

mpirun -np [num of processes] ./gpu-simulator.x 
  --configs [/path/to/application/configs] 
  --kernel_id [kernel id you want to evaluate] 
  --config_file ./DEV-Def/QV100.config

The options:

  • -np <num>: Number of processes.
  • --configs <path>: Specify the configuration file of the application
  • --kernel_id <num>: Specify the kernel ID for simulation

Example:

mpirun -np 10 ./gpu_simulator --configs ./apps/Rodinia/hotspot/configs --kernel_id 0 --config_file ./DEV-Def/QV100.config

Here, the ./apps/Rodinia/hotspot/configs are generated by tracing_tool. For detailed instructions, please refer to here.

After this, the simulator will generate output files (like kernel-0-rank-x.temp.txt, where x represents the process number) of many processes to the ./apps/Rodinia/hotspot/outputs directory.

Next, we need to merge the results of these processes using the script in the root directory of our simulator:

python merge_report.py --dir ./apps/Rodinia/hotspot/outputs --kernel_id 0 --np 10

The merged output file is named kernel-0-summary.txt and will be located in the ./apps/Rodinia/hotspot/outputs directory, and the content is roughly as follows:

 - Config: ./apps/Rodinia/hotspot/outputs

 - Theoretical Performance: 
       * Thread_block_limit_SM: 32
       * Thread_block_limit_registers: 6
       * Thread_block_limit_shared_memory: 32
       * Thread_block_limit_warps: 8
       * Theoretical_max_active_warps_per_SM: 48
       * Theoretical_occupancy: 0.75

 - L1 Cache Performance: 
       * Unified_L1_cache_hit_rate: 0.0349
       * Unified_L1_cache_requests: 216240

 - L2 Cache Performance: 
       * L2_cache_hit_rate: 0.6657
       * L2_cache_requests: 1186944
       * L2_read_transactions: 166020
       * L2_write_transactions: 42680
       * L2_total_transactions: 208700

 - DRAM Performance: 
       * DRAM_total_transactions: 49152

 - Global Memory Performance: 
       * GMEM_read_requests: 29240
       * GMEM_write_requests: 11008
       * GMEM_total_requests: 40248
       * GMEM_read_transactions: 172720
       * GMEM_write_transactions: 43520
       * GMEM_total_transactions: 216240
       * Number_of_read_transactions_per_read_requests: 5.91
       * Number_of_write_transactions_per_write_requests: 3.95
       * Total_number_of_global_atomic_requests: 0
       * Total_number_of_global_reduction_requests: 0
       * Global_memory_atomic_and_reduction_transactions: 0

 - SMs Performance: 
       * Achieved_active_warps_per_SM: 37.93
       * Achieved_occupancy: 0.75

       * GPU_active_cycles: 26087
       * SM_active_cycles: 26087

       * Warp_instructions_executed: 3073296
       * Instructions_executed_per_clock_cycle_IPC: 1.47
       * Total_instructions_executed_per_seconds (MIPS): 2182.06

       * Kernel_execution_time (ns): 18028

 - Simulator Performance: 
       * Simulation_time_memory_model (s): 0.37
       * Simulation_time_compute_model (s): 1.30

 - Stall Cycles: 
       * Compute_Structural_Stall: 6153
       * Compute_Data_Stall: 899
       * Memory_Structural_Stall: 3023
       * Memory_Data_Stall: 3501
       * Synchronization_Stall: 749
       * Control_Stall: 0
       * Idle_Stall: 90
       * Other_Stall: 0
       * No_Stall: 5489

 - Stall Cycles Distribution: 
       * Compute_Structural_Stall_ratio: 0.309134
       * Compute_Data_Stall_ratio: 0.045167
       * Memory_Structural_Stall_ratio: 0.151879
       * Memory_Data_Stall_ratio: 0.175894
       * Synchronization_Stall_ratio: 0.037631
       * Control_Stall_ratio: 0.000000
       * Idle_Stall_ratio: 0.004522
       * Other_Stall_ratio: 0.000000
       * No_Stall_ratio: 0.275774

 - Memory Structural Stall Cycles Breakdown: 
       * Issue_out_has_no_free_slot: 3023.0
       * Issue_previous_issued_inst_exec_type_is_memory: 0.0
       * Execute_result_bus_has_no_slot_for_latency: 0.0
       * Execute_m_dispatch_reg_of_fu_is_not_empty: 2010.0
       * Writeback_bank_of_reg_is_not_idle: 35.0
       * ReadOperands_bank_reg_belonged_to_was_allocated: 9.0
       * ReadOperands_port_num_m_in_ports_m_in_fails_as_not_found_free_cu: 2783.0
       * Execute_icnt_injection_buffer_is_full: 0.0

 - Memory Structural Stall Cycles Breakdown Distribution: 
       * Issue_out_has_no_free_slot: 0.384606
       * Issue_previous_issued_inst_exec_type_is_memory: 0.000000
       * Execute_result_bus_has_no_slot_for_latency: 0.000000
       * Execute_m_dispatch_reg_of_fu_is_not_empty: 0.255725
       * Writeback_bank_of_reg_is_not_idle: 0.004453
       * ReadOperands_bank_reg_belonged_to_was_allocated: 0.001145
       * ReadOperands_port_num_m_in_ports_m_in_fails_as_not_found_free_cu: 0.354071
       * Execute_icnt_injection_buffer_is_full: 0.000000

 - Compute Structural Stall Cycles Breakdown: 
       * Issue_out_has_no_free_slot: 6153.0
       * Issue_previous_issued_inst_exec_type_is_compute: 0.0
       * Execute_result_bus_has_no_slot_for_latency: 4.0
       * Execute_m_dispatch_reg_of_fu_is_not_empty: 5623.0
       * Writeback_bank_of_reg_is_not_idle: 518.0
       * ReadOperands_bank_reg_belonged_to_was_allocated: 927.0
       * ReadOperands_port_num_m_in_ports_m_in_fails_as_not_found_free_cu: 5240.0

 - Compute Structural Stall Cycles Breakdown Distribution: 
       * Issue_out_has_no_free_slot: 0.333225
       * Issue_previous_issued_inst_exec_type_is_compute: 0.000000
       * Execute_result_bus_has_no_slot_for_latency: 0.000217
       * Execute_m_dispatch_reg_of_fu_is_not_empty: 0.304522
       * Writeback_bank_of_reg_is_not_idle: 0.028053
       * ReadOperands_bank_reg_belonged_to_was_allocated: 0.050203
       * ReadOperands_port_num_m_in_ports_m_in_fails_as_not_found_free_cu: 0.283780

 - Memory Data Stall Cycles Breakdown: 
       * Issue_scoreboard: 3501.0
       * Execute_L1: 1337.0
       * Execute_L2: 95308.0
       * Execute_Main_Memory: 45181.0

 - Memory Data Stall Cycles Breakdown Distribution: 
       * Issue_scoreboard: 0.024090
       * Execute_L1: 0.009200
       * Execute_L2: 0.655818
       * Execute_Main_Memory: 0.310892

 - Compute Data Stall Cycles Breakdown: 
       * Issue_scoreboard: 899.0

 - Compute Data Stall Cycles Breakdown Distribution: 
       * Issue_scoreboard: 1.000000

 - Function Unit Execution Cycles Breakdown: 
       * SP_UNIT_execute_clks: 86027.0
       * SFU_UNIT_execute_clks: 50285.0
       * INT_UNIT_execute_clks: 765709.0
       * DP_UNIT_execute_clks: 82461.0
       * TENSOR_CORE_UNIT_execute_clks: 0.0
       * LDST_UNIT_execute_clks: 318340.0
       * SPEC_UNIT_1_execute_clks: 263957.0
       * SPEC_UNIT_2_execute_clks: 0.0
       * SPEC_UNIT_3_execute_clks: 0.0
       * Other_UNIT_execute_clks: 0.0

 - Function Unit Execution Cycles Breakdown Distribution: 
       * SP_UNIT_execute_clks: 0.054907
       * SFU_UNIT_execute_clks: 0.032095
       * INT_UNIT_execute_clks: 0.488715
       * DP_UNIT_execute_clks: 0.052631
       * TENSOR_CORE_UNIT_execute_clks: 0.000000
       * LDST_UNIT_execute_clks: 0.203181
       * SPEC_UNIT_1_execute_clks: 0.168471
       * SPEC_UNIT_2_execute_clks: 0.000000
       * SPEC_UNIT_3_execute_clks: 0.000000
       * Other_UNIT_execute_clks: 0.000000

 - Function Unit Execution Instns Number: 
       * SP_UNIT_Instns_num: 4004.0
       * SFU_UNIT_Instns_num: 736.0
       * INT_UNIT_Instns_num: 24267.0
       * DP_UNIT_Instns_num: 2177.0
       * TENSOR_CORE_UNIT_Instns_num: 0.0
       * LDST_UNIT_Instns_num: 3525.0
       * SPEC_UNIT_1_Instns_num: 3255.0
       * SPEC_UNIT_2_Instns_num: 0.0
       * SPEC_UNIT_3_Instns_num: 0.0
       * Other_UNIT_Instns_num: 0.0

 - Function Unit Execution Instns Number Distribution: 
       * SP_UNIT_Instns_num: 0.105468
       * SFU_UNIT_Instns_num: 0.019387
       * INT_UNIT_Instns_num: 0.639211
       * DP_UNIT_Instns_num: 0.057344
       * TENSOR_CORE_UNIT_Instns_num: 0.000000
       * LDST_UNIT_Instns_num: 0.092851
       * SPEC_UNIT_1_Instns_num: 0.085739
       * SPEC_UNIT_2_Instns_num: 0.000000
       * SPEC_UNIT_3_Instns_num: 0.000000
       * Other_UNIT_Instns_num: 0.000000

 - Function Unit Execution Average Cycles Per Instn : 
       * SP_UNIT_average_cycles_per_instn: 21.485265
       * SFU_UNIT_average_cycles_per_instn: 68.322011
       * INT_UNIT_average_cycles_per_instn: 31.553509
       * DP_UNIT_average_cycles_per_instn: 37.878273
       * TENSOR_CORE_UNIT_average_cycles_per_instn: 0.000000
       * LDST_UNIT_average_cycles_per_instn: 90.309220
       * SPEC_UNIT_1_average_cycles_per_instn: 81.092780
       * SPEC_UNIT_2_average_cycles_per_instn: 0.000000
       * SPEC_UNIT_3_average_cycles_per_instn: 0.000000
       * Other_UNIT_average_cycles_per_instn: 0.000000

 - Report Time: 2024-04-13 18:12:11

Docker

We are creating a docker image that will be released once AE is complete. You can preview it from Zenodo and the docker image is called micro57.tar.gz.