博客

1. 使用gem5 生成 simpoint
   cmdline：

build/X86/gem5.fast configs/spec2017/se_spec17.py --spec-2017-bench 
--spec-size=ref --warmup-insts=1000000 --maxinsts=100000000000 
--arch=X86 --cpu-type=AtomicSimpleCPU --sys-clock=1.6GHz 
--cpu-clock=3.5GHz --mem-size=16GB --mem-type=DDR4_2400_4x16 
--simpoint-profile 
--simpoint-interval 100000000  --l4-size=512MB -b mcf 

生成的文件目录，其中simpoint.bb.gz生成的压缩文件，供simpoint使用。

./mcf/
├── benchmark_err
├── benchmark_out
├── config.ini
├── config.json
├── dumpDebugState.txt
├── simerr
├── simout
├── simpoint.bb.gz
└── stats.txt

2. 使用simpoint处理simpoint.bb.gz 生成weights和simpoint
   见simpoint weight 格式
   此时我们得到了

./mcf/
├── benchmark_err
├── benchmark_out
├── config.ini
├── config.json
├── dumpDebugState.txt
├── simerr
├── simout
├── simpoint.bb.gz
├── simpoints
├── stats.txt
└── weights

3. gem5输入simpoints和weights，在对应的节点产生checkpoint
   我的实验warm up inst 4Billion比较长，可以根据自己需要修改，比如10K。

build/X86/gem5.fast 
configs/spec2017/se_spec17.py --spec-2017-bench --spec-size=ref 
--warmup-insts=1000000 --maxinsts=100000000000 
--arch=X86 --cpu-type=X86TimingSimpleCPU 
--sys-clock=1.6GHz --cpu-clock=3.5GHz --mem-size=16GB --mem-type=DDR4_2400_4x16 
--caches --cache-level=4 --l4-size=512MB 
-b mcf 
--take-simpoint-checkpoint=mcf/simpoints, mcf/weights,100000000,4000000000

4）产生了checkpoint之后的目录如下图所示：
这里我换了一个program，因为正好在跑这个。

cactuBSSN/
├── benchmark_err
├── benchmark_out
├── config.ini
├── config.json
├── cpt.simpoint_00_inst_6600000000_weight_0.177177_interval_100000000_warmup_4000000000
│   ├── m5.cpt
│   └── system.physmem.store0.pmem
├── cpt.simpoint_01_inst_49400000000_weight_0.058058_interval_100000000_warmup_4000000000
│   ├── m5.cpt
│   └── system.physmem.store0.pmem
├── cpt.simpoint_02_inst_51400000000_weight_0.014014_interval_100000000_warmup_4000000000
│   ├── m5.cpt
│   └── system.physmem.store0.pmem
├── cpt.simpoint_03_inst_56700000000_weight_0.612613_interval_100000000_warmup_4000000000
│   ├── m5.cpt
│   └── system.physmem.store0.pmem
├── cpt.simpoint_04_inst_82500000000_weight_0.138138_interval_100000000_warmup_4000000000
│   ├── m5.cpt
│   └── system.physmem.store0.pmem
├── simerr
├── simout
└── stats.txt

m5.cpt即为当前cpu运行状态的快照，比如寄存器值，页表映射关系之类的：

system.physmem.store0.pmem则是当时的内存的状态。
因此有了cpt和pmem，gem5就可以从任意一个checkpoint restore，开始执行。

5）restore checkpoint

build/X86/gem5.fast configs/spec2017/se_spec17.py
--spec-2017-bench --spec-size=ref 
--warmup-insts=1000000
--arch=X86 --cpu-type=X86O3CPU 
--sys-clock=1.6GHz --cpu-clock=3.5GHz --caches --cache-level=4 
--mem-size=16GB --mem-type=DDR4_2400_4x16 
--l4-size=256MB 
-b cactuBSSN
 --restore-simpoint-checkpoint -r 4 
 --checkpoint-dir gem5-results-2017-20220902-checkpoint/512MB/cactuBSSN

这里的 -r 4 比较特殊，这里计数是按1开头的，因此-r 4实际上对应 第4）步中的第3个checkpoint：
cpt.simpoint_03_inst_56700000000_weight_0.612613_interval_100000000_warmup_4000000000

Resuming from cactuBSSN/cpt.simpoint_03_inst_56700000000_weight_0.612613_interval_100000000_warmup_4000000000
Resuming from SimPoint #3, start_inst:56700000000, weight:0.612613, interval:100000000, warmup:4000000000
Global frequency set at 1000000000000 ticks per second
warn: No dot file generated. Please install pydot to generate the dot file and pdf.
build/X86/mem/dram_interface.cc:692: warn: DRAM device capacity (4096 Mbytes) does not match the address range assigned (16384 Mbytes)
0: system.remote_gdb: listening for remote gdb on port 7000
build/X86/sim/process.cc:405: warn: Checkpoints for pipes, device drivers and sockets do not work.
Switch at curTick count:10000
build/X86/sim/simulate.cc:194: info: Entering event queue @ 187482902990720.  Starting simulation...
Switched CPUS @ tick 187482903000720
switching cpus
system.cpu old->new system.switch_cpus
build/X86/sim/simulate.cc:194: info: Entering event queue @ 187482903000720.  Starting simulation...
build/X86/sim/power_state.cc:106: warn: PowerState: Already in the requested power state, request ignored
build/X86/sim/simulate.cc:194: info: Entering event queue @ 187482903000822.  Starting simulation...
O3CPU0 At 187487215508450 Tid[0] 10000000 instructions are executed.
O3CPU0 At 187491429812258 Tid[0] 20000000 instructions are executed.
O3CPU0 At 187495683554894 Tid[0] 30000000 instructions are executed.

使用restore checkpoint的方法就可以快速的从各个截点恢复执行，原理有点儿像分支预测里面的checkpoint和上下文切换的context。