Thermometer

Thermometer is a profile-guided BTB replacement technique.

Please cite the following paper if you use this artifact:

@inproceedings{song-thermometer-isca-2022,
  author = {Song, Shixin and Khan, Tanvir Ahmed and Mahdizadeh Shahri, Sara and Sriraman, Akshitha and Soundararajan, Niranjan K and Subramoney, Sreenivas and Jim{\'e}nez, Daniel A and Litz, Heiner and Kasikci, Baris},
  title = {Thermometer: Profile-Guided BTB Replacement for Data Center Applications},
  booktitle = {Proceedings of the 49th International Symposium on Computer Architecture (ISCA)},
  series = {ISCA 2022},
  year = {2022},
  video = {https://www.youtube.com/watch?v=YMqFlP0PtWw},
  code = {https://github.com/efeslab/thermometer-artifact},
  month = jun,
}

Simulator and trace

Thermometer is implemented in a trace-based simulator, ChampSim. Clone the Thermometer repo with

git clone git@github.com:efeslab/thermometer-artifact.git

We use Intel Processor Trace to generate traces for widely-used data center applications. The traces can be downloaded from here.

Update: We also released traces of four more apps here

Compile

Dependency

First, you should make sure that a C++ compiler (with C++17 support), cmake, and boost-filesystem are installed. Python 3 is required to install the dependency below, as well as to run the scripts.

The project also requires intelxed as a dependency. You can install it by

$ git clone https://github.com/intelxed/xed.git xed
$ git clone https://github.com/intelxed/mbuild.git mbuild
$ cd xed
$ python3 ./mfile.py --shared install

This will create a kits directory containing the compiled xed (both static and shared) libraries.

Add the absolute path to kits/xed-install-base-xxx to your env variable CMAKE_PREFIX_PATH so that CMake can find this package.

Build (Debug)

CMake will compile in debug mode by default. You can use this build to debug your code.

$ mkdir cmake-build-debug
$ cd cmake-build-debug
$ cmake ..
$ make -j

Build (Release)

You can also add a flag to compile in the release mode. This is useful in the real experiments since the code is optimized for speed.

$ mkdir cmake-build-release
$ cd cmake-build-release
$ cmake -DCMAKE_BUILD_TYPE=Release ..
$ make -j

Experiments

Compiled executables

After successfully build all executables, you can find them in the directory cmake-build-debug or cmake-build-release. Different executables are corresponding to different experimental setup mentioned in the paper. We list corresponding names of major experiment setups in the following tables. For some experimental setups, the corresponding executable names can be found according to the experiment running scripts.

Executable	BTB Replacement	BTB Prefetch	Other Setup
ChampSim_fdip_lru	LRU
ChampSim_fdip_hot_warm_cold_80_50_f_keep_curr_hotter_lru	Thermometer
ChampSim_fdip_opt	Optimal Replacement
ChampSim_fdip_srrip	SRRIP
ChampSim_fdip_ghrp	GHRP
ChampSim_fdip_hawkeye	Hawkeye
ChampSim_fdip_predecoder_btb_lru	LRU	Confluence
ChampSim_fdip_shotgun_lru	LRU	Shotgun
ChampSim_fdip_predecoder_btb_hot_warm_cold_80_50_f_keep_curr_hotter_lru	Thermometer	Confluence
ChampSim_fdip_shotgun_hot_warm_cold_80_50_f_keep_curr_hotter_lru	Thermometer	Shotgun
ChampSim_fdip_predecoder_btb_opt	Optimal Replacement	Confluence
ChampSim_fdip_shotgun_opt	Optimal Replacement	Shotgun
ChampSim_fdip_perfect_btb			No BTB Misses
ChampSim_fdip_perfect_bp	LRU		Correct branch direction
ChampSim_icache_lru			No I-Cache Misses (very large I-Cache)

Run Experiments

Use the following script to run most experiments

run_test.py

Use the following script to generate result summary from raw data

python3 generate_result.py --trace-type=pt --total-btb-ways=4 --dir-list=ChampSim_fdip_lru,ChampSim_fdip_srrip,ChampSim_fdip_ghrp,ChampSim_fdip_hawkeye,ChampSim_fdip_opt,ChampSim_fdip_hot_warm_cold_80_50_f_keep_curr_hotter_lru,ChampSim_fdip_perfect_btb,ChampSim_fdip_perfect_bp,ChampSim_icache_lru,ChampSim_fdip_predecoder_btb_lru,ChampSim_fdip_shotgun_lru,ChampSim_fdip_predecoder_btb_hot_warm_cold_80_50_f_keep_curr_hotter_lru,ChampSim_fdip_shotgun_hot_warm_cold_80_50_f_keep_curr_hotter_lru,ChampSim_fdip_predecoder_btb_opt,ChampSim_fdip_shotgun_opt,7979ChampSim_fdip_opt,7979ChampSim_fdip_hot_warm_cold_80_50_f_keep_curr_hotter_lru

Run Experiments and Generate Results

Since some Figures use the data from the same experiment, so please run the following commands in order to get required raw data to generate the final results.

IPC Speedup and BTB Miss Reduction

First, run experiments using the script run_test.py.

# LRU, SRRIP, GHRP, Hawkeye
# Perfect BTB, Perfect Branch Prediction, Perfect I-Cache
# LRU + Confluence, LRU + Shotgun
python3 run_test.py --trace-type=pt --total-btb-ways=4 --run-list=ChampSim_fdip_lru,ChampSim_fdip_srrip,ChampSim_fdip_ghrp,ChampSim_fdip_hawkeye,ChampSim_fdip_perfect_btb,ChampSim_fdip_perfect_bp,ChampSim_icache_lru,ChampSim_fdip_predecoder_btb_lru,ChampSim_fdip_shotgun_lru --app-list=cassandra,clang,drupal,finagle-chirper,finagle-http,kafka,mediawiki,mysqllarge,pgbench,python,tomcat,verilator,wordpress
# Optimal Replacement,
# Optimal Replacement + Confluence, Optimal Replacement + Shotgun
python3 run_test.py --trace-type=pt --generate=True --total-btb-ways=4 --run-list=ChampSim_fdip_opt,ChampSim_fdip_predecoder_btb_opt,ChampSim_fdip_shotgun_opt --app-list=cassandra,clang,drupal,finagle-chirper,finagle-http,kafka,mediawiki,mysqllarge,pgbench,python,tomcat,verilator,wordpress
# Thermometer, Thermometer (holistic information only)
# Thermometer + Confluence, Thermometer + Shotgun
python3 run_test.py --trace-type=pt --total-btb-ways=4 --run-list=ChampSim_fdip_hot_warm_cold_80_50_f_keep_curr_hotter_lru,ChampSim_fdip_hwc_50_80_f_keep_curr_hotter,ChampSim_fdip_predecoder_btb_hot_warm_cold_80_50_f_keep_curr_hotter_lru,ChampSim_fdip_shotgun_hot_warm_cold_80_50_f_keep_curr_hotter_lru --app-list=cassandra,clang,drupal,finagle-chirper,finagle-http,kafka,mediawiki,mysqllarge,pgbench,python,tomcat,verilator,wordpress
# Optimal Replacement and Thermometer on BTB with 7979 entry
python3 run_test.py --trace-type=pt --generate=True --total-btb-ways=4 --total-btb-entries=7979 --run-list=ChampSim_fdip_opt --app-list=cassandra,clang,drupal,finagle-chirper,finagle-http,kafka,mediawiki,mysqllarge,pgbench,python,tomcat,verilator,wordpress
python3 run_test.py --trace-type=pt --total-btb-ways=4 --total-btb-entries=7979 --run-list=ChampSim_fdip_hot_warm_cold_80_50_f_keep_curr_hotter_lru --app-list=cassandra,clang,drupal,finagle-chirper,finagle-http,kafka,mediawiki,mysqllarge,pgbench,python,tomcat,verilator,wordpress
# Twig
#python3 run_test.py --trace-type=pt --total-btb-ways=4 --total-btb-entries=8 --run-list=ChampSim_fdip_lru --twig-profile=True --app-list=cassandra,clang,drupal,finagle-chirper,finagle-http,kafka,mediawiki,mysqllarge,pgbench,python,tomcat,verilator,wordpress
## TODO: One step of generating Twig result is missing
#python3 run_test.py --trace-type=pt --total-btb-ways=4 --total-btb-entries=8 --run-list=ChampSim_fdip_lru,ChampSim_fdip_srrip --use-twig-prefetch=True
#python3 run_test.py --trace-type=pt --total-btb-ways=4 --total-btb-entries=8 --run-list=ChampSim_fdip_opt --use-twig-prefetch=True --app-list=cassandra,clang,drupal,finagle-chirper,finagle-http,kafka,mediawiki,mysqllarge,pgbench,python,tomcat,verilator,wordpress
#python3 run_test.py --trace-type=pt --total-btb-ways=4 --total-btb-entries=8 --run-list=ChampSim_fdip_hot_warm_cold_80_50_f_keep_curr_hotter_lru --use-twig-prefetch=True --app-list=cassandra,clang,drupal,finagle-chirper,finagle-http,kafka,mediawiki,mysqllarge,pgbench,python,tomcat,verilator,wordpress

Second, generate result summary using the script generate_result.py.

python3 generate_result.py --trace-type=pt --total-btb-ways=4 --dir-list=ChampSim_fdip_lru,ChampSim_fdip_srrip,ChampSim_fdip_ghrp,ChampSim_fdip_hawkeye,ChampSim_fdip_opt,ChampSim_fdip_hot_warm_cold_80_50_f_keep_curr_hotter_lru,ChampSim_fdip_hwc_50_80_f_keep_curr_hotter,ChampSim_fdip_perfect_btb,ChampSim_fdip_perfect_bp,ChampSim_icache_lru,ChampSim_fdip_predecoder_btb_lru,ChampSim_fdip_shotgun_lru,ChampSim_fdip_predecoder_btb_opt,ChampSim_fdip_shotgun_opt,ChampSim_fdip_predecoder_btb_hot_warm_cold_80_50_f_keep_curr_hotter_lru,ChampSim_fdip_shotgun_hot_warm_cold_80_50_f_keep_curr_hotter_lru,7979ChampSim_fdip_opt,7979ChampSim_fdip_hot_warm_cold_80_50_f_keep_curr_hotter_lru

Then, calculate corresponding IPC speedup and generate plots. You can check comments in the script to find out map from figure number to file name and plot code.

# Fig. 1 3 4 11
python3 plots/scripts/plot_ipc_mpki.py
# Fig. 2
python3 plots/scripts/plot_functions.py

Finally, calculate and plot BTB miss reduction.

# Fig. 12
python3 plots/scripts/plot_btb_miss.py

Reuse Distance Variance

# Fig. 5
python3 plots/scripts/plot_reuse_distance_variance.py

Hit-to-taken percentage

# Fig. 6 7
python3 plots/scripts/plot_hit_access_count.py

Correlation

# Fig. 8
python3 plots/scripts/plot_hwc_bias_corr.py

Bypass

# Fig. 9
python3 plots/scripts/plot_opt_eviction.py

Coverage

# Fig. 9
python3 plots/scripts/plot_coverage.py

Accuracy

# Fig. 9
python3 plots/scripts/plot_accuracy.py

Note

Some paths are hard-coded in the implementation and scripts, feel free to replace them with what you needed. Most result data and plots are already included in this repo. If you have any problems (e.g., scripts do not work. some results cannot be generated), feel free to contact us.