Usage of incorrect scaling factor for latency measurement

[geethaprasuna29]

Hello all,

I reproduced the DRAMgym setup at my end and observed a huge jump in latency numbers (up to 3 orders of magnitude) when one of the parameters is tweaked slightly. I noticed that this is because of using a common scaling factor (1e9) while reporting the observations vector [energy, power, latency] from the DRAMSys simulator.

design point 1:
{'PagePolicy': 'OpenAdaptive', 'Scheduler': 'FrFcfs', 'SchedulerBuffer': 'Shared', 'RequestBufferSize': 1, 'RespQueue': 'Reorder', 'RefreshPolicy': 'NoRefresh', 'RefreshMaxPostponed': 2, 'RefreshMaxPulledin': 2, 'Arbiter': 'Reorder', 'MaxActiveTransactions': 1}

Actual: Total Energy: 432139050.00 pJ', 'DRAMSys.dram0 Average Power: 1002.89 mW', 'DRAMSys.controller0 Total Time: 430895 ns',
Reported: [Environment] Observation: [4.3213905e-01 1.0028900e+00 4.3089500e-04]

design point 2:
{'PagePolicy': 'OpenAdaptive', 'Scheduler': 'FrFcfs', 'SchedulerBuffer': 'Shared', 'RequestBufferSize': 1, 'RespQueue': 'Reorder', 'RefreshPolicy': 'NoRefresh', 'RefreshMaxPostponed': 2, 'RefreshMaxPulledin': 2, 'Arbiter': 'Reorder', 'MaxActiveTransactions': 2}

Actual: Total Energy: 354977775.00 pJ', 'DRAMSys.dram0 Average Power: 1232.55 mW', 'DRAMSys.controller0 Total Time: 288003750 ps',
Reported: [Environment] Observation: [3.54977775e-01 1.23255000e+00 2.88003750e-01]

Though the units are ps, using the same scaling factor 1e9 is causing the latency to be reported as 0.288 sec when it is actually 0.000288 sec

Relevant Code:
line number 66 in oss-arch-gym/arch_gym/envs/DRAMEnv.py

  def get_observation(self,outstream):
       '''
       converts the std out from DRAMSys to observation of energy, power, latency
       [Energy (PJ), Power (mW), Latency (ns)]
       '''
       obs = []
       
       keywords = ["Total Energy", "Average Power", "Total Time"]

       energy = re.findall(keywords[0],outstream)
       all_lines = outstream.splitlines()
       for each_idx in range(len(all_lines)):
           
           if keywords[0] in all_lines[each_idx]:
               obs.append(float(all_lines[each_idx].split(":")[1].split()[0])/1e9)
           if keywords[1] in all_lines[each_idx]:
               obs.append(float(all_lines[each_idx].split(":")[1].split()[0])/1e3)
           if keywords[2] in all_lines[each_idx]:
               **obs.append(float(all_lines[each_idx].split(":")[1].split()[0])/1e9)**
       
       obs = np.asarray(obs)
       print('[Environment] Observation:', obs)
       
       if(len(obs)==0):
            print(outstream)
       return obs