- About PIMSim
PIMSim is a trace-based simulator of PIM architecture, which provides both host-side and memory-side simulation. Besides, PIMSim provides high degree of freedom of PIM unit configuration which can be defined by users. PIMSim provides three input types to adapt different simulation demanding. PIMSim aims to support new memory types such as HMC, HBM et.
- What can PIMSim do?
PIMSim provides simulation of Process-In-Memory(PIM) and applies to the following groups:
+ People who want to get insight of PIM architecture.
+ Researchers who want to build new PIM designs.
+ Programmers who want to discover the potential of PIM architecture.
PIMSim can provide following experimental statistics:
+ PIM
--Speed up while applying new PIM designs
--Off-Chip Bandwidth
--Message statistics
--Internal bandwidth
--Energy (under development)
--PIMProcessor simulation details
--Pipeline simulation details
--Data Coherence of PIM architecture
--Program partitions details
+ Processor
--Cache simulation details
--MSHR simulation details
--ALU simulation details
--Hybrid memory simulation details
To take a deep understanding of PIMSim, we’ll show some cases to take advantage of PIMSim:
Case1: Develop new PIM design. PIMSim is a simulator that simulates host-side CPU, memory, and PIM units (processors or computational logic) at the same time. You can implement your design logic in PIMSim and it will bring you detailed simulation statistics.
Case2: Test programs and explore the potential of PIM architecture. If you want to know the performance of your code on PIM architecture, or test your program partitions for best performance, you can use PIMSim to partition your programs and easily get your performance results.
Case3: Get insight of PIM architecture. If you want to know the internal bandwidth of PIM or the detailed off-chip traffic, you can use PIMSim to get this information.
- Getting started with PIMSim
3.1 Prepare input traces.
PIMSim is a trace-based simulator. All the trace filename should be “ CPUx.trace ” (x indicate the id of CPU, or you can customize your own inputs). Currently, PIMSim supports two types of traces: Detailed trace input and PC-based trace input. Detailed trace input is aim to get detailed simulation of instructions and cycles. PC-based trace input is designed for best simulation speed and the situations that only cares about memory behaviors. The traces can be fetched by either running programs on full-system simulator or dynamic program analysis tools on physical machines. The formats of trace line are listed below:
3.1.1 Trace Formation
PC-Based trace input:
[64bits PC] [64bits memory access address(first bit 0: Read, first bit 1: Write]
7f8a8af36ff6 7f8a8b154b80
7f8a8af36ffd 80007f8a8b154cd8
7f8a8af37007 7f8a8b154fb8
Detailed trace format：
[cycle]|[instruction]|[read/write]|[data]|[address]
0|ld t1, SS:[rsp]|R|D=0x0000000000000001 A=0x7de145ffee20
1|addi rsp, rsp, 0x8
2|ld t1, SS:[rsp]|R|D=0x0000000000000001 A=0x024ee20
4|mov rdx, rdx, rsp
5|limm t1, 0xfffffffffffffff0
6|and rsp, rsp, t1
7|st rax, SS:[rsp + 0xffdfff8]|W|D=0x0000000000000000 A=0x04fee18
3.1.2 PIM Partition methods
In PC-Based trace input, you can customize your PIM kernal PC (start pc and end pc) in config files.
In Detailed trace input, there are four input labels are applied to indicate target operation is executed at memory-side:
a. You can use “PIM_Operation” to indicate that this instruction should be executed in memory-side units, like this :
26|PIM_st r14, SS:[rsp + 0xffffff8]|W|D=0x0000000000000000 A=0x163df5
b. If you want to nominate a snippet of code to be executed in memory-side, you can label traces like this:
PIM_BLOCK_START
18|rdip t7, %ctrl153,
19|st t7, SS:[rsp + 0xffffffffff8]|W|D=0x00000000004001ba A=0x7fffffffee08
20|subi rsp, rsp, 0x8
21|wrip , t7, t1
22|st r15, SS:[rsp + 0xfffffffff8]|W|D=0x0000000000000000 A=0x7fffffffee00
PIM_BLOCK_END
c. If you don’t know the actual instructions executed but know the memory access address, we provide function to help. You can add this to use function:
3|Function_Add_Start ; cycle/function name
input = 0x7fffffffee20 ;input address
input = 0x7fffffffee18 ; input address
output = 0x5807b0 ;output address
latency = 2 ;operation latency without data fetching
Function_Add_End
This input indicates an function named 'Add', which has two inputs and one output. The operation duration lasts 2 cycles.
d. You can customize your own PIM operation policy by modifying PIMConfigs.cs. For example, you can specify a certain kind of instruction to be executed at memory-side by modifying PIMConfig\PIM_Ins_List.
3.2 Prepare the configuration files
PIMSim has two kinds of configuration inputs: PIM settings and RAM settings. For PIM settings, you can attach them by PIMConfig Class in "/Configs/PIMConfigs.cs". For RAM settings, we modified and integrate HMCSim (https://github.com/tactcomplabs/gc64-hmcsim) and DRAMSim2 (https://github.com/dramninjasUMD/DRAMSim2) to adapt PIM Simulation. If you need simulate HMC or DRAM (or both of them), you should provide configure files of them. Detailed document of HMCSIM and DRAMSim2 can be found on their GitHub. In the future, we'll add more new memory simulations such as NVM, Memristor and so on.
3.3 How to get Detailed PIM kernal addresses?
You can use three methods to get detailed PIM kernal information:
+ Instruction Instrumentation Tool
Pintool (https://software.intel.com/en-us/articles/pin-a-dynamic-binary-instrumentation-tool)
+ Performance Profiler
OProfile (http://oprofile.sourceforge.net/news/)
VTune (https://software.intel.com/en-us/intel-vtune-amplifier-xe/)
+ Hardware Profiler
HMTT (http://asg.ict.ac.cn/hmtt/)
3.4 Building PIMSim
To build PIMSim, you can locate the root folder of PIMSim and type:
$make
To clean the previous buildings, you can type:
$make clean
3.5 Running PIMSim
You can run PIMSim by providing such paramaters:
$PIMSim -t tracefilepath -config configfilepath –o outputfile –n processorcount –c cycle
-t, -trace FILEPATH specify the path folder of input trace.
-config FILEPATH specify the path folder of input configs.
-o, -output FILENAME specify the file name of output file."
-n, -N PROCCOUNT specify the count of host proc."
-c, -cycle CYCLES specify the execution cycles.