HMCTherm

HMCTherm: A Cycle-accurate Simulator for Hybrid Memory Cube with thermal analysis version HMCTherm v1.0 - 2018.04.25

1. Developer

Zhiyuan Yang
Ankur Srivastava
University of Maryland
zyyang [at] umd [dot] edu

2. About HMCTherm

HMCTherm is a comprehensive simulation framework for Stacked-Memory-on-CPU architecture. Given the architectural description of multi-core CPUs and HMCs, HMCTherm can simulate the 3D thermal profile (both transient and static) of the HMC when a certain program is running. The HMCTherm is built based on CasHMC v1.2 (https://github.com/estwings57/CasHMC) which takes memory traces as inputs and performs cycle-accurate simulation of HMC to get the delay, power and thermal information.

A grid-based thermal solver is integrated to HMCTherm which is based on SuperLU_MT (http://crd-legacy.lbl.gov/~xiaoye/SuperLU/). SuperLU is a general purpose library for the direct solution of large, sparse, nonsysmmetric systems of linear equations; and SuperLU_MT is a version that enables multi-threaded computation.

3. Folder Directory

sources : source files for the HMC simulator
Example : containing memory traces for the simulation
script : containing python scripts for ploting final power and thermal profiles

4. Usage

Dependency

The integrated grid-based thermal solver is based on SuperLU_MT. SuperLU_MT should be installed before installing HMCTemp. SuperLU_MT can be downloaded from http://crd-legacy.lbl.gov/~xiaoye/SuperLU/ and following the instruction to install the library.

Installation

(1) install the architectural simulator

$ cd multi2sim_memtrace
$ make

(2) install McPAT

$ cd McPAT
$ make

(3) in the Makefile for the HMC simulator (under the root directory of HMCTherm), modify the SuperLUroot to where the SuperLU is installed in your local system
(4) install the HMC simulator

$ make

Run your Simulation

(1) run the HMC simulator

$ cd $(HMCThermROOT)/
$ ./HMCTherm -c 400000 -t file -f ./Example/SPEC_CPU2006_example/mase_trace_bwaves_base.alpha.v0.trc -x 1024 -y 1024 -e 20000 -d ./

In this command:
-c indicates the number of CPU cycles to be simulated
-t indicates whether the input memory trace is from a file (file) or generated randomly (random)
-f if the memory trace is from a file, the trace file name is specified here
-x and -y specifies the size (in byte) of a "mat" which is the smallest unit of memory [Default 512]
-e is the time step to print out the transient temperature and power profile
-d specifies the directory saving the output files
-h will print out the help of the HMC simulator

The HMC simulation will generate five output files:
- /result/file.out : summary of the HMC latency
- temperature_trace.csv : the temperature traces for each time step
- power_trace.csv : the power traces for each time step
- Average_power.csv : the average power profile within the time of simulation
- static_temperature.csv : the static thermal profile given the average power

(2) Plot the power and temperature profile using the python scripts in $(HMCThermROOT)/script/
- plot_trans.py : plot the transient power or temperature profile
- plot_staticT.py : plot the static temperature profile
- plot_staticP.py : plot the average power profile

akothen / HMCTherm