PMSI: Predictable MSI cache coherence protocol
This is the gem5 code repository for Predictable MSI (PMSI) presented in our RTAS'17 paper: Predictable Cache Coherence for Multi-core Real-time Systems.
- The Gem5 simulator has been modified to support predictable snoopy bus architecture, and trace based simulation.
$GEM5refers to the top level directory where gem5 resides.
$GEM5/build_opts/PMSIcontains information for building the PMSI system, and
$GEM5/configs/ruby/PMSI.pyprovides information about the component (caches, network, DRAM) connections in the system. Note that these files do not need to be changed unless the cache hierarchy is extended.
- To build PMSI, execute
scons $GEM5/build/PMSI/gem5.opt -j8. The
-j8is to accelerate the build process.
Predictable bus arbitration
- We extend Gem5 to support a snoopy bus with a predictable arbiter. The predictable aribiter operates in a time division multiplexed (TDM) manner.
src/cpu/testers/rubyTest/Trace.hhcontrol the number of requestors and the TDM slot width per requestor.
Memory trace simulation
- Gem5 has been extended to run memory trace based execution.
- The trace based simulation injects traces into the ruby memory model thereby bypassing the core/processor.
- Trace based simulation can be enabled by setting the
- The trace based simulation reads from a file
trace.trcthat consists of lines of requests of the form
Addr OP time. OP is of type RD for read and WR for write, and time is a positive integer value that denotes the arrival time of the memory request to the memory hierarchy.
- The trace simulation automatically replicates the memory operations in
trace.trcacross multiple cores (specified by
-nflag) with the same time-stamp. The issue of the request to the bus is handled by the TDM arbiter. To see a complete trace of the memory requests issued by the cores, enable the protocol trace by using the following command line option:
- A sample trace file can be found in
- To run PMSI, the following command line options need to be used for correct functioning:
--ruby, --cpu-type=timing, --cpu-clock=xGHz --ruby-clock=yGHz.
- Other options such as cache sizes, memory size, cache associativity, and number of cores can be configured accordingly.
- For binary/normal execution, supply the binary name and the associated arguments using the
-oflags respectively. For example, to run FFT of the Splash2 benchmark:
./build/X86_MSI_Snooping_One_Level_RT/gem5.opt --stats-file=FFT-pmsi-stats.out ./configs/example/se.py --ruby -n4 --cpu-type=timing --l1d_size=16kB --l1i_size=16kB --l1d_assoc=1 --l1i_assoc=1 --mem-size=4194304kB --topology=Crossbar --cpu-clock=2GHz --ruby-clock=2GHz --mem-type=SimpleMemory -c $SPLASH/splash2/codes/kernels/fft/fft -o "-m 10 -p 4 -t"
- For trace based execution, populate the
trace.trcfile with the necessary memory operations and execute:
./build/X86_MSI_Snooping_One_Level_RT/gem5.opt ./configs/example/ruby_random_test.py --ruby-clock=2GHz --ruby --cpu-clock=2GHz --topology=Crossbar --mem-type=SimpleMemory -n 4 --mem-size=4194304kB --wakeup_freq=1.
- We provide a script to run the SPLASH2 benchmarks in the
- We also provide a docker environment to replicate the environment we used to carry out our evaluation. Please run
sudo docker build -t pmsi .to build the container with dependencies and PMSI, and
sudo docker run -i -t pmsito run.
- Feel free to contact us for questions regarding PMSI.