2cke_transOvlp



In [1]:
%load_ext autoreload
%autoreload 2

In [2]:
import warnings
import pandas as pd
import numpy as np
import os
import sys # error msg, add the modules
import operator # sorting
from math import *
import matplotlib.pyplot as plt

sys.path.append('../../')

import read_trace
import cuda_timeline
# from avgblkmodel import *
from ModelParam import *
import cke
from df_util import *
#from model_cke import *

warnings.filterwarnings("ignore", category=np.VisibleDeprecationWarning)

gpu info


In [3]:
gtx950 = DeviceInfo()
gtx950.sm_num = 6
gtx950.sharedmem_per_sm = 49152
gtx950.reg_per_sm = 65536
gtx950.maxthreads_per_sm = 2048

2 stream info


In [4]:
# 10M for mem_mem : where the h2d between streams are overlapped
trace_file = 'trace_10M_s1.csv'
trace_file_2cke = 'trace_h2d_h2d_ovlp.csv'

df_trace = read_trace.trace2dataframe(trace_file) # read the trace to the dataframe
df_trace_2cke = read_trace.trace2dataframe(trace_file_2cke)

In [5]:
#df_trace

In [6]:
#cuda_timeline.plot_trace(df_trace)

In [7]:
df_trace_2cke


Out[7]:
Start Duration Grid X Grid Y Grid Z Block X Block Y Block Z Registers Per Thread Static SMem Dynamic SMem Size Throughput Device Context Stream Name
0 ms ms NaN NaN NaN NaN NaN NaN NaN B B MB GB/s NaN NaN NaN NaN
1 665.032627 10.783099 NaN NaN NaN NaN NaN NaN NaN NaN NaN 38.146973 3.454749 GeForce GTX 950 (0) 1.0 13.0 [CUDA memcpy HtoD]
2 668.190628 12.898855 NaN NaN NaN NaN NaN NaN NaN NaN NaN 38.146973 2.888078 GeForce GTX 950 (0) 1.0 14.0 [CUDA memcpy HtoD]
3 675.816942 12.913127 NaN NaN NaN NaN NaN NaN NaN NaN NaN 38.146973 2.884886 GeForce GTX 950 (0) 1.0 13.0 [CUDA memcpy HtoD]
4 681.090667 10.213144 NaN NaN NaN NaN NaN NaN NaN NaN NaN 38.146973 3.647545 GeForce GTX 950 (0) 1.0 14.0 [CUDA memcpy HtoD]
5 688.741397 1.377896 39063.0 1.0 1.0 256.0 1.0 1.0 8.0 0 0 NaN NaN GeForce GTX 950 (0) 1.0 13.0 kernel_vectorAdd(float const *, float const *,...
6 690.126461 10.658490 NaN NaN NaN NaN NaN NaN NaN NaN NaN 38.146973 3.495139 GeForce GTX 950 (0) 1.0 13.0 [CUDA memcpy DtoH]
7 691.316867 1.366120 39063.0 1.0 1.0 256.0 1.0 1.0 8.0 0 0 NaN NaN GeForce GTX 950 (0) 1.0 14.0 kernel_vectorAdd(float const *, float const *,...
8 693.294222 9.522868 NaN NaN NaN NaN NaN NaN NaN NaN NaN 38.146973 3.911942 GeForce GTX 950 (0) 1.0 14.0 [CUDA memcpy DtoH]

In [8]:
cuda_timeline.plot_trace(df_trace_2cke)


/home/leiming/anaconda2/lib/python2.7/site-packages/matplotlib/axes/_base.py:1292: UnicodeWarning: Unicode equal comparison failed to convert both arguments to Unicode - interpreting them as being unequal
  if aspect == 'normal':
/home/leiming/anaconda2/lib/python2.7/site-packages/matplotlib/axes/_base.py:1297: UnicodeWarning: Unicode equal comparison failed to convert both arguments to Unicode - interpreting them as being unequal
  elif aspect in ('equal', 'auto'):

1cke - read trace and reset the timeline


In [9]:
df_single_stream = read_trace.get_timing(df_trace)

In [10]:
df_single_stream


Out[10]:
stream api_type start end size duration
0 0.0 h2d 610.840271 617.277086 38146.973 6.436815
1 0.0 h2d 617.278302 623.708301 38146.973 6.429999
2 0.0 kern 623.719789 624.996407 0.000 1.276618
3 0.0 d2h 625.003191 631.272837 38146.973 6.269646

In [11]:
df_s1 = read_trace.reset_starting(df_single_stream)

In [12]:
df_s1


Out[12]:
stream api_type start end size duration
0 0.0 h2d 0.000000 6.436815 38146.973 6.436815
1 0.0 h2d 6.438031 12.868030 38146.973 6.429999
2 0.0 kern 12.879518 14.156136 0.000 1.276618
3 0.0 d2h 14.162920 20.432566 38146.973 6.269646

2cke case


In [13]:
df_2stream = read_trace.get_timing(df_trace_2cke)

In [14]:
df_2stream


Out[14]:
stream api_type start end size duration
0 0.0 h2d 665.032627 675.815726 38146.973 10.783099
1 0.0 h2d 675.816942 688.730069 38146.973 12.913127
2 0.0 kern 688.741397 690.119293 0.000 1.377896
3 0.0 d2h 690.126461 700.784951 38146.973 10.658490
4 1.0 h2d 668.190628 681.089483 38146.973 12.898855
5 1.0 h2d 681.090667 691.303811 38146.973 10.213144
6 1.0 kern 691.316867 692.682987 0.000 1.366120
7 1.0 d2h 693.294222 702.817090 38146.973 9.522868

In [15]:
tot_runtime = read_trace.getTotalRuntime(df_2stream)
print tot_runtime


37.784463

2 cke


In [16]:
stream_num = 2

# find when to start the stream and update the starting pos for the trace
H2D_H2D_OVLP_TH = 3.158431

df_cke_list = cke.init_trace_list(df_s1, stream_num = stream_num, h2d_ovlp_th = H2D_H2D_OVLP_TH)

In [17]:
df_cke_list[0]


Out[17]:
stream api_type start end size duration
0 0 h2d 0.000000 6.436815 38146.973 6.436815
1 0 h2d 6.438031 12.868030 38146.973 6.429999
2 0 kern 12.879518 14.156136 0.000 1.276618
3 0 d2h 14.162920 20.432566 38146.973 6.269646

In [18]:
df_cke_list[1]


Out[18]:
stream api_type start end size duration
0 1 h2d 3.158431 9.595246 38146.973 6.436815
1 1 h2d 9.596462 16.026461 38146.973 6.429999
2 1 kern 16.037949 17.314567 0.000 1.276618
3 1 d2h 17.321351 23.590997 38146.973 6.269646

sort


In [19]:
df_all_api = cke.init_sort_api_with_extra_cols(df_cke_list)

In [20]:
df_all_api


Out[20]:
start end api_type size_kb stream_id status bw bytes_done bytes_left current_pos time_left pred_end
0 0.000000 6.436815 h2d 38146.973 0.0 sleep 5926.373991 0.0 38146.973 0.0 0.0 0.0
4 3.158431 9.595246 h2d 38146.973 1.0 sleep 5926.373991 0.0 38146.973 0.0 0.0 0.0
1 6.438031 12.868030 h2d 38146.973 0.0 sleep 5932.656133 0.0 38146.973 0.0 0.0 0.0
5 9.596462 16.026461 h2d 38146.973 1.0 sleep 5932.656133 0.0 38146.973 0.0 0.0 0.0
2 12.879518 14.156136 kern 0.000 0.0 sleep 0.000000 0.0 0.000 0.0 0.0 0.0
3 14.162920 20.432566 d2h 38146.973 0.0 sleep 6084.390251 0.0 38146.973 0.0 0.0 0.0
6 16.037949 17.314567 kern 0.000 1.0 sleep 0.000000 0.0 0.000 0.0 0.0 0.0
7 17.321351 23.590997 d2h 38146.973 1.0 sleep 6084.390251 0.0 38146.973 0.0 0.0 0.0

start algo


In [21]:
count = 1
# break_count = 7

while not cke.AllDone(df_all_api):
    # pick two api to learn 
    df_all_api, r1, r2 = cke.PickTwo(df_all_api)
    
    if r1 == None and r2 == None:                          # go directly updating the last wake api
        df_all_api = cke.UpdateStream_lastapi(df_all_api)
    else:
        df_all_api = cke.StartNext_byType(df_all_api, [r1, r2])

        whichType = cke.CheckType(df_all_api, r1, r2) # check whether the same api
#         print whichType

        if whichType == None:
            df_all_api = cke.Predict_noConflict(df_all_api, r1, r2)
        elif whichType in ['h2d', 'd2h']: # data transfer in the same direction
            df_all_api = cke.Predict_transferOvlp(df_all_api, r1, r2, ways = 2.0)
        else: # concurrent kernel: todo
            pass

    #     if count == break_count:
    #         break

        rangeT = cke.Get_pred_range(df_all_api)
#         print rangeT

    #     if count == break_count:
    #         break

        extra_conc = cke.Check_cc_by_time(df_all_api, rangeT) # check whether there is conc during the rangeT

        if extra_conc == 0:
            if whichType in ['h2d', 'd2h']:
                df_all_api = cke.Update_wake_transferOvlp(df_all_api, rangeT, ways = 2.0)
            elif whichType == 'kern':
                pass
            else: # no overlapping
                df_all_api = cke.Update_wake_noConflict(df_all_api, rangeT)

            # check if any api is done, and update the timing for the other apis in that stream
            df_all_api = cke.UpdateStreamTime(df_all_api)

        else: # todo : when there is additional overlapping
            pass

#         if count == break_count:
#             break
        
    # next call
    count = count + 1

In [22]:
df_all_api


Out[22]:
start end api_type size_kb stream_id status bw bytes_done bytes_left current_pos time_left pred_end
0 0.000000 9.715199 h2d 38146.973 0.0 done 5926.373991 38146.973 0.0 9.715199 0.0 9.715199
4 3.158431 16.030845 h2d 38146.973 1.0 done 5926.373991 38146.973 0.0 16.030845 0.0 16.030845
1 9.716415 22.575197 h2d 38146.973 0.0 done 5932.656133 38146.973 0.0 22.575197 0.0 22.575197
5 16.032061 25.733628 h2d 38146.973 1.0 done 5932.656133 38146.973 0.0 25.733628 0.0 25.733628
2 22.586685 23.863303 kern 0.000 0.0 done 0.000000 0.000 0.0 23.863303 0.0 23.863303
3 23.870087 33.250948 d2h 38146.973 0.0 done 6084.390251 38146.973 0.0 33.250948 0.0 33.250948
6 25.745116 27.021734 kern 0.000 1.0 done 0.000000 0.000 0.0 27.021734 0.0 27.021734
7 27.028518 36.409379 d2h 38146.973 1.0 done 6084.390251 38146.973 0.0 36.409379 0.0 36.409379

In [23]:
df_all_api.loc[df_all_api.stream_id == 0]


Out[23]:
start end api_type size_kb stream_id status bw bytes_done bytes_left current_pos time_left pred_end
0 0.000000 9.715199 h2d 38146.973 0.0 done 5926.373991 38146.973 0.0 9.715199 0.0 9.715199
1 9.716415 22.575197 h2d 38146.973 0.0 done 5932.656133 38146.973 0.0 22.575197 0.0 22.575197
2 22.586685 23.863303 kern 0.000 0.0 done 0.000000 0.000 0.0 23.863303 0.0 23.863303
3 23.870087 33.250948 d2h 38146.973 0.0 done 6084.390251 38146.973 0.0 33.250948 0.0 33.250948

In [24]:
df_all_api.loc[df_all_api.stream_id == 1]


Out[24]:
start end api_type size_kb stream_id status bw bytes_done bytes_left current_pos time_left pred_end
4 3.158431 16.030845 h2d 38146.973 1.0 done 5926.373991 38146.973 0.0 16.030845 0.0 16.030845
5 16.032061 25.733628 h2d 38146.973 1.0 done 5932.656133 38146.973 0.0 25.733628 0.0 25.733628
6 25.745116 27.021734 kern 0.000 1.0 done 0.000000 0.000 0.0 27.021734 0.0 27.021734
7 27.028518 36.409379 d2h 38146.973 1.0 done 6084.390251 38146.973 0.0 36.409379 0.0 36.409379

In [25]:
#
# run above
#