hoangt/chainsaw

Chainsaw-Sim: A Chain simulator

Chainsaw-Sim is a cycle accurate simulator that models the host core, the Chainsaw accelerator, and spatial fabrics of parameterizable size.

Citation

	@inproceedings{sharifian2016chainsaw,
		title={Chainsaw: Von-neumann accelerators to leverage fused instruction chains},
		author={Sharifian, Amirali and Kumar, Snehasish and Guha, Apala and Shriraman, Arrvindh},
		booktitle={Microarchitecture (MICRO), 2016 49th Annual IEEE/ACM International Symposium on},
		pages={1--14},
		year={2016},
		organization={IEEE}
	}

Dependencies

• 1- Boost 1.61 (or higher)
• 2- CMake 2.8.8 (or higher)
• 3- C++11 compiler (e.g., clang++, g++-5)

Build

1. Download boost: You can download boost library from here

   Note: If you already have built boost on your system you can skip first step

   Compile boost with regex, graph, system, filesystem, iostreams and program_option libraries.

    $ ./bootstrap --show-libraries (showing available boost libraries)
    $ ./bootstrap.sh --prefix=`<PATH_TO_INSTALL>` --with-libraries=filesystem,graph,regex,program_options,system
    $ ./b2 --prefix=`<PATH_TO_INSTALL>` install
   

2. Clone the tools from the Chainsaw-Sim GitHub repository:

    $ git clone [email protected]:sfu-arch/chainsaw.git
   

3. Compile: Simulator with CMake:

    $ cd ChainSaw-Sim
    $ mkdir build && cd build
    $ cmake ../ -DBOOST_ROOT=<BOOST_ROOT_DIR>/boost_1_61_0/ -DCMAKE_BUILD_TYPE=Release
    $ make && make install
   

4. Test: the simulator

For running Chainsaw you need to add these two paths into your LD_LIBRARY_PATH.

$ export <BOOST_ROOT_DIR>/lib:${LD_LIBRARY_PATH}
$ export <CHAINSAW_SRC>/gems-lib/ruby_clean/amd64-linux/generated/MESI_CMP_directory_m/lib:${LD_LIBRARY_PATH}

$ make test

If you could pass the test case you will the following message:

Running tests...
    Start 1: Test
    1/1 Test #1: Test .............................   **Passe**    9.27 sec

    100% tests passed, 0 tests failed out of 1

    Total Test time (real) =   9.28 sec

In case you couldn't pass the test case you will get the following message:

Running tests...
    Start 1: Test
    1/1 Test #1: Test .............................***Failed    0.00 sec

    0% tests passed, 1 tests failed out of 1

    Total Test time (real) =   0.01 sec

    The following tests FAILED:
          1 - Test (Failed)
          Errors while running CTest

The following file contains error message:

$ cat <BUILD_DIR>/Testing/Temporary/LastTest.log

Usage

In def folder there are three config file which set the simulator's variables:

1. Graph.def:

Graph.def contains configurations for input graph.

• <MAX_ILP>: Breaking each chain at every live-in and live-out so that we can get highest available ILP from the chains.
• <MERGE>: Running mergeq algorithm on input graph.
• <FUSE>: Running memory fusion mode.
• <MAX_NUM_CHAIN>: Maximum number chains.
• <MAX_LIVE_OUT>: Maximum number of live-out for each chain.
• <MAX_LIVE_IN>: Maximum number of live-in for each chain.
• <BLOCK_LIMIT>: Number of bypass registers.

2. Core.def:

Core.def contains Core's config variables. By setting these variables you can control the Core's fabric parameters:

• <MEMORY_MODE>: Choosing core's memory system.
• <LANE_SIZE>: Set how many instructions each lane can support.
• <SCHEDULER>: Chainsaw supports two different scheduling strategies.
• <NUM_LANE>: Set number of lanes.

3. Lane.def

• <L1_MISS_LATENCY>: L1 cache miss latency.
• <L2_MISS_LATENCY>: L2 cache miss latency.
• <L3_MISS_LATENCY>: L3 cache miss latency.
• <FETCH_DELAY>: Fetch stage's delay.
• <DECODE_DELAY>: Decode stage's delay.

Getting Started

Chainsaw in general runs in two different modes:

1. Memory Trace Driven: In this mode Chainsaw directly reads memory from a file, and simulate the Memory subsystem while it's simulating the instructions. There two different memory files which should feed to Chainsaw simulator. In the first file each line contains only memory requests within the hotpath and also node id inside the data flow graph and iteration id (Format: <ITER_ID>,<Load|Store>,<NODE_ID>,<ADDRESS>).

    .
    .
    2,Load,10,140683921899632
    2,Store,14,140683921899608
    .
    .
   

The second contains program memory requests and also markers indicating beginning of the HotPath trace:

    .
    .
    140721745227768,Store,8
    140721745227736,Store,8
    2,__chunk_begin_
    140721745227736,Load,8
    140721745227776,Load,8
    140721745227784,Load,8
    .
    .

The addresses have been saved in decimal format and each node ID should match with the node ID inside original dot graph file.

Running Chainsaw in Memory trace driven mode:

    $ cd build
    $ make -j install
    $ ./bin/chainSaw -i <DOT_GRAPH> -a <HOTPATH_TRACE> -m <MEMORY_TRACE>
    $ Simulation is done.
    # NOTE: Statics they are written in output/ folder.

2. Constant memory mode In this mode the user has three options to put constant miss-latency delay for all the memory requests:

Running Chainsaw in Memory trace driven mode:

    $ cd build
    $ make -j install
    $ ./bin/chainSaw -i <DOT_GRAPH> 
    $ Simulation is done.
    # NOTE: Statics they are written in output/ folder.

These modes can be set in the config files.

Input file

Chainsaw uses dataflow graph in order to simulate the hot path. The chainer project works as a preprocessing step in order to prepare Chainsaw input files. We rely on Git's submodule system to take care of resolving the references. Chainer has been added to the Chainsaw simulator as a submodule in order to fetch the project you need to run:

$ cd Chainsaw-Sim
$ git submodule update --init --recursive

chainer has instructions how to run it on input dot graph files.

Simulation output

In the output folder you can find the simulation statics:

• <name-chain.dot>: Chain's graph file.
• <name-chainOpHisto.csv>: Histogram of instructions in different chains with different size.
• <name-chainSize.csv>: Histogram of the chain sizes.
• <name-chainStat.csv>: Final chain's graph stats.
• <name-chILP.dot>: Chain ILP at each level.
• <name-ilpgraph.dot>: Loss ILP graph.
• <name-graphHisto.csv>: Input graph operations histogram.
• <name-graphStat.csv>: Input graph stats.
• <name-liveInHisto.csv>: Live-in histogram for chains.
• <name-liveOutHisto.csv>: Live-out histogram for chains.

Contributors

• Amirali Sharifian (Simon Fraser University)
• Apala Guha (Simon Fraser University)
• Snehasish Kumar (Simon Fraser University)