项目介绍

umd-memsys/DRAMSim2 是一个托管在 GitHub 上的开源项目,主要使用 C++ 编写,目前获得 306 个 Star、166 次 Fork。

仓库全名
umd-memsys/DRAMSim2
Star 数
306
Fork 数
166
未关闭 Issue
33

README 摘要

DRAMSim 1. About DRAMSim --------------------------------------------------------- DRAMSim is a cycle accurate model of a DRAM memory controller, the DRAM modules which comprise system storage, and the bus by which they communicate. All major components in a modern memory system are modeled as their own respective objects within the source, including: ranks, banks, command queue, the memory controller, etc. We also…

完整内容见 GitHub 原始 README 或本页 。

常见问题

DRAMSim2 用什么语言开发?

主要使用 C++ 编写,目前有 306 个 Star。

如何获取 DRAMSim2 的源码?

可以直接克隆仓库:git clone https://github.com/umd-memsys/DRAMSim2.git,默认分支为 master;安装与运行步骤见仓库 README。

DRAMSim2 最近还在维护吗?

仓库最近一次更新为 2020年11月12日,当前有 33 个未关闭 Issue;Star 数为 306,可作为社区活跃度的参考。

关于

GitHub 原始描述(英文)

DRAMSim2: A cycle accurate DRAM simulator

描述取自 GitHub 官方 API,保持仓库作者书写的原文。

简要分析

热度分
666
价值分
234
活跃状态
活跃

README

DRAMSim

  1. About DRAMSim --------------------------------------------------------- DRAMSim is a cycle accurate model of a DRAM memory controller, the DRAM modules which comprise system storage, and the bus by which they communicate. All major components in a modern memory system are modeled as their own respective objects within the source, including: ranks, banks, command queue, the memory controller, etc.

    We also added support and config files for STT-MRAM, you can run the simulations the same way as DRAM. Refer Asifuzzaman, K., etc. Enabling a reliable STT-MRAM main memory simulation (MEMSYS2017) for more info about STT-MRAM.

    The overarching goal is to have a simulator that is extremely small, portable, and accurate. The simulator core has a well-defined interface which allows it to be CPU simulator agnostic and should be easily modifiably to work with any simulator. This core has no external run time or build time dependencies and has been tested with g++ on Linux as well as g++ on Cygwin on Windows.

  2. Building DRAMSim ---------------------------------------------------------

    To build an optimized DRAMSim simply type

    $ make

    For a debug build which contains debugging symbols and verbose output, run:

    $ make DEBUG=1

    this will compile an executable called DRAMSim which can run a trace-based simulation.

    To build the DRAMSim library, type:

    $ make libdramsim.so

  3. Running DRAMSim -----------------------------------------------------------

    First, run the preprocessor on the gzipped traces:

    cd traces ./traceParse.py k6_aoe_02_short.trc.gz

    Then go back to the DRAMSim directory and run the trace based simulator: cd … ./DRAMSim -t traces/k6_aoe_02_short.trc -s system.ini -d ini/DDR3_micron_64M_8B_x4_sg15.ini -c 10000

  4. DRAMSim Output -------------------------------------------------------------

The verbosity of the DRAMSim can be customized in the ini file by turning the various debug flags on or off in the ini file.

Below, we have provided a detailed explanation of the simulator output. With all DEBUG flags enabled, the following output is displayed for each cycle executed.

NOTE : BP = Bus Packet T = Transaction MC = MemoryController R# = Rank (index #)


----------------- Memory System Update ------------------ ---------- Memory Controller Update Starting ------------ [8] – R0 Receiving On Bus : BP [ACT] pa[0x5dec7f0] r[0] b[3] row[1502] col[799] – MC Issuing On Data Bus : BP [DATA] pa[0x7edc7e0] r[0] b[2] row[2029] col[799] data[0]= ++ Adding Read energy to total energy – MC Issuing On Command Bus : BP [READ_P] pa[0x5dec7f8] r[1] b[3] row[1502] col[799] == New Transaction - Mapping Address [0x5dec800] Rank : 0 Bank : 0 Row : 1502 Col : 800 ++ Adding IDD3N to total energy [from rank 0] ++ Adding IDD3N to total energy [from rank 1] == Printing transaction queue 8]T [Read] [0x45bbfa4] 9]T [Write] [0x55fbfa0] [5439E] 10]T [Write] [0x55fbfa8] [1111] == Printing bank states (According to MC) [idle] [idle] [2029] [1502] [idle] [idle] [1502] [1502]

== Printing Per Rank, Per Bank Queue = Rank 0 Bank 0 size : 2 0]BP [ACT] pa[0x5dec800] r[0] b[0] row[1502] col[800] 1]BP [READ_P] pa[0x5dec800] r[0] b[0] row[1502] col[800] Bank 1 size : 2 0]BP [ACT] pa[0x5dec810] r[0] b[1] row[1502] col[800] 1]BP [READ_P] pa[0x5dec810] r[0] b[1] row[1502] col[800] Bank 2 size : 2 0]BP [ACT] pa[0x5dec7e0] r[0] b[2] row[1502] col[799] 1]BP [READ_P] pa[0x5dec7e0] r[0] b[2] row[1502] col[799] Bank 3 size : 1 0]BP [READ_P] pa[0x5dec7f0] r[0] b[3] row[1502] col[799] = Rank 1 Bank 0 size : 2 0]BP [ACT] pa[0x5dec808] r[1] b[0] row[1502] col[800] 1]BP [READ_P] pa[0x5dec808] r[1] b[0] row[1502] col[800] Bank 1 size : 2 0]BP [ACT] pa[0x5dec818] r[1] b[1] row[1502] col[800] 1]BP [READ_P] pa[0x5dec818] r[1] b[1] row[1502] col[800] Bank 2 size : 1 0]BP [READ_P] pa[0x5dec7e8] r[1] b[2] row[1502] col[799] Bank 3 size : 0

Anything sent on the bus is encapsulated in an BusPacket (BP) object. When printing, they display the following information: BP [ACT] pa[0x5dec818] r[1] b[1] row[1502] col[800] The information displayed is (in order):command type, physical address, rank #, bank#, row #, and column #.

Lines beginning with " – " indicate bus traffic, ie, – R0 Receiving On Bus : BP [ACT] pa[0x5dec7f0] r[0] b[3] row[1502] col[799] – MC Issuing On Data Bus : BP [DATA] pa[0x7edc7e0] r[0] b[2] row[2029] col[799] data[0]= – MC Issuing On Command Bus : BP [READ_P] pa[0x5dec7f8] r[1] b[3] row[1502] col[799] Sender and receiver are indicated and the packet being sent is detailed.

Lines beginning with " ++ " indicate power calculations, ie, ++ Adding Read energy to total energy ++ Adding IDD3N to total energy [from rank 0] ++ Adding IDD3N to total energy [from rank 1] The state of the system and the actions taken determine which current draw is used. for further detail about each current, see micron data- sheet.

If a pending transaction is in the transaction queue, it will

be printed, as seen below: == Printing transaction queue 1]T [Read] [0x45bbfa4] 2]T [Write] [0x55fbfa0] [5439E] 3]T [Write] [0x55fbfa8] [1111] Currently, at the start of every cycle, the head of the transaction queue is removed, broken up into DRAM commands and placed in the appropriate command queues. To do this, an address mapping scheme is applied to the transaction’s physical address, the output of which is seen below: == New Transaction - Mapping Address [0x5dec800] Rank : 0 Bank : 0 Row : 1502 Col : 800

If there are pending commands in the command queue, they will be printed. The output is dependent on the designated structure for the command queue. For example, per-rank/per-bank queues are shown below: = Rank 1 Bank 0 size : 2 0]BP [ACT] pa[0x5dec808] r[1] b[0] row[1502] col[800] 1]BP [READ_P] pa[0x5dec808] r[1] b[0] row[1502] col[800] Bank 1 size : 2 0]BP [ACT] pa[0x5dec818] r[1] b[1] row[1502] col[800] 1]BP [READ_P] pa[0x5dec818] r[1] b[1] row[1502] col[800] Bank 2 size : 1 0]BP [READ_P] pa[0x5dec7e8] r[1] b[2] row[1502] col[799] Bank 3 size : 0

The state of each bank in the system is also displayed: == Printing bank states (According to MC) [idle] [idle] [2029] [1502] [idle] [idle] [1502] [1502] Banks can be in many states, including idle, row active (shown with the row that is active), refreshing, or precharging. These states will update based on the commands being sent by the controller.

评论

暂无评论