def main():
parser = argparse.ArgumentParser()
parser.add_argument('csv', type=str, help='Full path to CSV file')
args = parser.parse_args()
c = args.csv
m = "*"
if not ".txt" in args.csv:
print("Suspicious input file with no/wrong extension:", args.csv)
exit(1)
e = Experiment(c)
ranges, faults, batches, num_batches = e.get_raw_data()
e.print_info()
matplotlib.rcParams['agg.path.chunksize'] = 10000
fig = plt.figure()
#fig = plt.gcf()
ax = fig.add_subplot(1, 1, 1)
faults = sorted(faults, key=lambda f: f.timestamp)
read = []
rx = []
write = []
wx = []
for i, f in enumerate(faults):
if f.access_type == "r":
read.append(f.fault_address)
rx.append(i)
elif f.access_type == "w":
write.append(f.fault_address)
wx.append(i)
else:
print("unaccounted for access type:", faults.access_type)
#print(ranges)
ax.hlines(ranges, 0, len(faults))
ax.plot(rx, read, 'b' + m, markersize=1, label="read")
ax.plot(wx, write, 'r' + m, markersize=1, label="write")
#fig.set_size_inches(16, 10)
plt.legend()
ax.set_xlabel('Fault Occurence')
ax.set_ylabel('Fault Fault Index')
psize = dirname(args.csv).split("_")[1]
figname = splitext(basename(
args.csv))[0].split("_")[0] + "-" + psize + "-time.png"
ylabels = map(lambda t: '0x%013X' % int(t), ax.get_yticks())
ax.set_yticklabels(ylabels)
plt.tight_layout()
print('saving figure:', figname)
fig.savefig(figname, dpi=500)
plt.close(fig)
def parse_bsize_sub(csv):
e = Experiment(csv)
size = None
if "pf" in csv:
size = basename(dirname(csv)).split("_")[2]
else:
size = basename(dirname(csv)).split("_")[1]
return (e, size)
def parse_bsize(t):
bsize, lis = t
x = []
y = []
for csv in humansorted(lis):
e = Experiment(csv)
size = None
total_dups = None
if "pf" in csv:
size = basename(dirname(csv)).split("_")[2]
total_dups = sum(e.get_duplicate_faults_64k())
else:
size = basename(dirname(csv)).split("_")[1]
total_dups = sum(e.get_duplicate_faults_4k())
size = size
x.append(size)
y.append(total_dups)
return (x, y, bsize)
def parse_bsize_sub2(t):
bsize, csv = t
e = Experiment(csv)
size = None
if "pf" in csv:
size = basename(dirname(csv)).split("_")[2]
else:
size = basename(dirname(csv)).split("_")[1]
app = splitext(basename(csv))[0].split("_")[0]
return ExperimentStats(e, size, bsize, app)
def per_app(c):
e = Experiment(c)
ranges, faults, batches, num_batches = e.get_raw_data()
e.print_info()
matplotlib.rcParams['agg.path.chunksize'] = 10000
x0, y0 = do_batch_len_time(batches, faults, c)
x1, y1 = do_batch_len_time_4k(batches, faults, c)
x2, y2 = do_batch_len_time_64k(batches, faults, c)
x3, y3 = do_batch_len_time_2m(batches, faults, c)
do_all(batches, faults, c, x0, x1, x2, x3, y0, y1, y2, y3)
x4, y4 = do_batch_len_time_transfers_4k(batches, faults, c)
x5, y5 = do_batch_len_time_transfers_64k(batches, faults, c)
do_all_more(batches, faults, c, x0, x1, x2, x3, x4, x5, y0, y1, y2, y3, y4,
y5)
print(c)
pages_per_vablock_per_batch(batches, faults)
def parse_bsize(t):
bsize, lis = t
es = []
for csv in humansorted(lis):
e = Experiment(csv)
size = None
if "pf" in csv:
size = basename(dirname(csv)).split("_")[2]
else:
size = basename(dirname(csv)).split("_")[1]
es.append(ExperimentStats(e, size, bsize))
return (es, bsize)
def main():
parser = argparse.ArgumentParser()
parser.add_argument('csv', type=str, help='Full path to CSV file')
args = parser.parse_args()
c = args.csv
m = "*"
if not ".txt" in args.csv:
print("Suspicious input file with no/wrong extension:", args.csv)
exit(1)
e = Experiment(c)
ranges, faults, batches, num_batches = e.get_raw_data()
e.print_info()
matplotlib.rcParams['agg.path.chunksize'] = 10000
fig = plt.figure()
#fig = plt.gcf()
ax = fig.add_subplot(1,1,1)
sm_ids = sorted(e.count_utlb_client_pairs().keys())
cmap = plt.get_cmap('jet')
colors = cmap(np.linspace(0, 1.0, len(sm_ids)))
id_f_map = {sm_id:[] for sm_id in sm_ids}
id_fx_map = {sm_id:[] for sm_id in sm_ids}
x = 1
for fault in faults:
mapping = (fault.utlb_id, fault.client_id)
id_f_map[mapping].append(fault.fault_address)
id_fx_map[mapping].append(x)
x += 1
#print(ranges)
ax.hlines(ranges, 0, len(faults))
for i, (sm_id, color) in enumerate(zip(sm_ids, colors)):
#ax.scatter(id_fx_map[sm_id], id_f_map[sm_id] , color=color, label=str(sm_id), s=.01)
ax.plot(id_fx_map[sm_id], id_f_map[sm_id] , color=color, label=str(sm_id), linestyle="None", markersize=1, marker=",")
#if i == 0:
# break
#fig.set_size_inches(16, 10)
#plt.legend()
ax.set_xlabel('Fault Occurence')
ax.set_ylabel('Fault Fault Index')
psize = dirname(args.csv).split("_")[1]
figname = splitext(basename(args.csv))[0].split("_")[0] + "-" + psize + "-utlb.png"
ylabels = map(lambda t: '0x%013X' % int(t), ax.get_yticks())
ax.set_yticklabels(ylabels)
plt.tight_layout()
print('saving figure:', figname)
fig.savefig(figname, dpi=800)
plt.close(fig)
def parse_bsize(t):
bsize, lis = t
x = []
y = []
for csv in humansorted(lis):
e = Experiment(csv)
size = None
total_dups = None
if "pf" in csv:
size = basename(dirname(csv)).split("_")[2]
else:
size = basename(dirname(csv)).split("_")[1]
size = size
x.append(size)
y.append(len(e.batches))
return (x, y, bsize)
def main():
parser = argparse.ArgumentParser()
parser.add_argument('csv', type=str, help='Full path to CSV file')
parser.add_argument('-o', type=str, default="", help='output filename')
args = parser.parse_args()
c = args.csv
m = "*"
if not ".txt" in args.csv:
print("Suspicious input file with no/wrong extension:", args.csv)
exit(1)
e = Experiment(c)
ranges, faults, batches, num_batches = e.get_raw_data()
e.print_info()
matplotlib.rcParams['agg.path.chunksize'] = 10000
fig = plt.figure()
#fig = plt.gcf()
#ax = fig.add_subplot(1,1,1)
#ax2 = ax.twiny()
#print(batches)
plt.clf()
ax = plt.gca()
x = []
y = []
for i, batch in enumerate(batches):
y_set = set()
for fault in batch:
if fault.utlb_id not in y_set:
x.append(i)
y_set.add(fault.utlb_id)
y += [y for y in y_set]
plt.plot(x, y, "bo", markersize=1)
print("max tlb:", max(y))
print("unique tlb", len(set(y)))
plt.xlabel("Batch ID")
plt.ylabel("UTLB Fault Present")
#plt.legend()
psize = basename(dirname(args.csv)) #.split("_")[-1]
print("psize:", psize)
figname = None
if args.o == "":
figname = (splitext(basename(args.csv))[0] + "-" + psize +
"-utlb-active.png").replace("_", "-")
else:
figname = args.o
if ".png" not in figname:
figname += ".png"
#figname = splitext(basename(args.csv))[0].split("_")[0] + "-" + psize + "-sm-time-dist.png"
plt.tight_layout()
print('saving figure:', figname)
fig.savefig(figname, dpi=800)
plt.close(fig)
def main():
parser = argparse.ArgumentParser()
parser.add_argument('csv', type=str, help='Full path to CSV file')
parser.add_argument('-o', type=str, default="", help='output filename')
args = parser.parse_args()
c = args.csv
m = "*"
if not ".txt" in args.csv:
print("Suspicious input file with no/wrong extension:", args.csv)
exit(1)
e = Experiment(c)
ranges, faults, batches, num_batches = e.get_raw_data()
e.print_info()
matplotlib.rcParams['agg.path.chunksize'] = 10000
fig = plt.figure()
#fig = plt.gcf()
#ax = fig.add_subplot(1,1,1)
#ax2 = ax.twiny()
read = []
rx = []
write = []
wx = []
for i, f in enumerate(faults):
if f.access_type == "r":
#read.append(f.fault_address)
#rx.append(i)
pass
elif f.access_type == "w":
#write.append(f.fault_address)
#wx.append(i)
pass
elif f.access_type == "p":
pass
else:
print("unaccounted for access type:", f.access_type)
true_bl = [len(batch) for batch in batches]
bl = [0] + true_bl
for i, b in enumerate(bl):
if i == 0:
continue
bl[i] += bl[i - 1]
del bl[0]
#print(batches)
sm_ids = sorted(e.count_utlb_client_pairs().keys())
cmap = plt.get_cmap('jet')
colors = cmap(np.linspace(0, 1.0, len(sm_ids)))
id_f_map = {sm_id: [] for sm_id in sm_ids}
for fault in faults:
mapping = (fault.utlb_id, fault.client_id)
id_f_map[mapping].append(fault)
times_between_faults = []
for sm_id in id_f_map.keys():
for first, second in zip(
sorted(id_f_map[sm_id][0:-1], key=lambda f: f.timestamp),
sorted(id_f_map[sm_id][1:], key=lambda f: f.timestamp)):
times_between_faults.append(second.timestamp - first.timestamp)
avg = np.mean(times_between_faults)
print("avg time between faults:", avg)
print("min time between faults", min(times_between_faults))
print("max time between faults", max(times_between_faults))
print("total # deltas:", len(times_between_faults))
print("Num faults < 1e1:",
len([t for t in times_between_faults if t < 1e1]))
print("Num faults < 1e2:",
len([t for t in times_between_faults if t < 1e2]))
print("Num faults < 1e3:",
len([t for t in times_between_faults if t < 1e3]))
print("Num faults < 1e4:",
len([t for t in times_between_faults if t < 1e4]))
print("Num faults < 1e5:",
len([t for t in times_between_faults if t < 1e5]))
print("Num faults < 1e6:",
len([t for t in times_between_faults if t < 1e6]))
print("Num faults < 1e7:",
len([t for t in times_between_faults if t < 1e7]))
print("Num faults < 1e8:",
len([t for t in times_between_faults if t < 1e8]))
Q1 = np.quantile(times_between_faults, 0.25)
Q3 = np.quantile(times_between_faults, 0.75)
med = statistics.median(times_between_faults)
print("Q1, median, Q3:", Q1, ",", med, ",", Q3)
#plt.plot(times, counts, marker="*")
hist, bins, _ = plt.hist(times_between_faults, bins=16)
print("bins:", bins)
logbins = np.logspace(0.0, np.log10(bins[-1]), len(bins))
print("logbins:", logbins)
plt.clf()
hist, bins, _ = plt.hist(times_between_faults, bins=logbins)
ax = plt.gca()
plt.vlines([Q1, med, Q3],
0,
1,
transform=ax.get_xaxis_transform(),
label="Q1/Med/Q3")
plt.xlim(xmin=1e0)
plt.ylim(ymin=0.0)
plt.xscale("log")
plt.xlabel("Time Between Same-SM Fault Arrival in Buffer (NS)")
plt.ylabel("Frequency")
plt.legend()
psize = basename(dirname(args.csv)) #.split("_")[-1]
print("psize:", psize)
figname = None
if args.o == "":
figname = (splitext(basename(args.csv))[0] + "-" + psize +
"-sm-time-dist.png").replace("_", "-")
else:
figname = args.o
if ".png" not in figname:
figname += ".png"
#figname = splitext(basename(args.csv))[0].split("_")[0] + "-" + psize + "-sm-time-dist.png"
plt.tight_layout()
print('saving figure:', figname)
fig.savefig(figname, dpi=500)
plt.close(fig)
def main():
parser = argparse.ArgumentParser()
parser.add_argument('csv', type=str, help='Full path to CSV file')
args = parser.parse_args()
c = args.csv
m = "*"
if not ".txt" in args.csv:
print("Suspicious input file with no/wrong extension:", args.csv)
exit(1)
e = Experiment(c)
ranges, faults, batches, num_batches = e.get_raw_data()
e.print_info()
matplotlib.rcParams['agg.path.chunksize'] = 10000
fig = plt.figure()
#fig = plt.gcf()
ax = fig.add_subplot(1, 1, 1)
#ax2 = ax.twiny()
read = []
rx = []
write = []
wx = []
prefetch = []
px = []
itot = 0
discard = []
dx = []
for batch, dbatch in zip(batches, e.dbatches):
batch = sorted(batch + dbatch, key=lambda f: f.timestamp)
for i, f in enumerate(batch):
if f.access_type == "r":
read.append(f.fault_address)
rx.append(i + itot)
elif f.access_type == "w":
write.append(f.fault_address)
wx.append(i + itot)
elif f.access_type == "p":
prefetch.append(f.fault_address)
px.append(i + itot)
elif f.access_type == "d":
discard.append(f.fault_address)
dx.append(i + itot)
else:
print("unaccounted for access type:", faults[i].access_type)
itot += len(batch)
for dbatch in e.dbatches:
batch = sorted(dbatch, key=lambda f: f.timestamp)
#for i,f in enumerate(sorted(faults, key=lambda f: f.timestamp)):
# if f.access_type == "r":
# read.append(f.fault_address)
# rx.append(i + itot)
# elif f.access_type == "w":
# write.append(f.fault_address)
# wx.append(i + itot)
# else:
# print("unaccounted for access type:", faults.access_type)
base_addr = min(ranges)
print("base_addr:", base_addr)
#base_addr = min(faults, key=lambda f: f.fault_address).fault_address
read = [(f - base_addr) // 4096 for f in read]
write = [(f - base_addr) // 4096 for f in write]
prefetch = [(f - base_addr) // 4096 for f in prefetch]
discard = [(f - base_addr) // 4096 for f in discard]
ranges = [(r - base_addr) // 4096 for r in ranges]
print(read)
print(write)
bl = [0] + [len(batch) for batch in batches]
for i, b in enumerate(bl):
if i == 0:
continue
bl[i] += bl[i - 1]
del bl[0]
#print(batches)
print("batch lengths:", [len(b) for b in batches])
print("batch indices:", bl)
#print(ranges)
ylim1 = min(read + write + prefetch + discard)
ylim2 = max(read + write + prefetch + discard)
ax.vlines(bl, min(ranges), ylim2, label="batch", color="k", linewidth=1)
ax.hlines(ranges,
0,
len(faults),
label="malloc",
linestyles="dotted",
color="k",
linewidth=1)
#ax.hlines(ranges, 0, len(faults), label="cudaMallocManaged()", linestyles="dotted", color="k", linewidth=1)
ax.plot(rx, read, 'b' + m, markersize=2, label="read")
ax.plot(wx, write, 'r' + m, markersize=2, label="write")
if (len(px) > 0):
ax.plot(px, prefetch, 'g' + m, markersize=2, label="prefetch")
if (len(dx) > 0):
ax.plot(dx, prefetch, 'g' + m, markersize=2, label="discard")
#newfaults = [f.fault_address for f in sorted(faults, key=lambda f: f.timestamp)]
#newx = [f.timestamp for f in sorted(faults, key=lambda f: f.timestamp)]
#newx = [int(f - min(newx)) for f in newx]
#ax2.plot(newx, newfaults, 'g'+m, markersize=1, label="time_order")
#print("newx", newx)
#fig.set_size_inches(16, 10)
ax.set_xlabel('Fault')
ax.set_ylabel('Page Index')
psize = dirname(args.csv).split("_")[1]
figname = splitext(basename(
args.csv))[0].split("_")[0] + "-" + psize + "-batch-time.png"
#ylabels = map(lambda t: '0x%013X' % int(t), ax.get_yticks())
#ax.set_yticklabels(ylabels)
ax.legend(loc="upper left")
#h1, l1 = ax.get_legend_handles_labels()
#h2, l2 = ax2.get_legend_handles_labels()
#ax.legend(h1+h2, l1+l2, loc="upper left")
#ax2.set_xlabel("Time (NS)");
plt.tight_layout()
print('saving figure:', figname)
fig.savefig(figname, dpi=500)
plt.close(fig)
def main():
parser = argparse.ArgumentParser()
parser.add_argument('csv', type=str, help='Full path to CSV file')
args = parser.parse_args()
c = args.csv
m = "*"
if not ".txt" in args.csv:
print("Suspicious input file with no/wrong extension:", args.csv)
exit(1)
e = Experiment(c)
ranges, faults, batches, num_batches = e.get_raw_data()
e.print_info()
matplotlib.rcParams['agg.path.chunksize'] = 10000
fig = plt.figure()
#fig = plt.gcf()
ax = fig.add_subplot(1,1,1)
ax2 = ax.twiny()
read=[]
rx = []
write=[]
wx = []
for i,f in enumerate(faults):
if f.access_type == "r":
read.append(f.fault_address)
rx.append(i)
elif f.access_type == "w":
write.append(f.fault_address)
wx.append(i)
else:
print("unaccounted for access type:", faults.access_type)
bl = [0] + [len(batch) for batch in batches]
for i,b in enumerate(bl):
if i == 0:
continue
bl[i] += bl[i-1]
del bl[0]
#print(batches)
print("batch lengths:", bl)
sm_ids = sorted(e.count_utlb_client_pairs().keys())
cmap = plt.get_cmap('jet')
colors = cmap(np.linspace(0, 1.0, len(sm_ids)))
id_f_map = {sm_id:[] for sm_id in sm_ids}
id_fx_map = {sm_id:[] for sm_id in sm_ids}
x = 1
for fault in faults:
mapping = (fault.utlb_id, fault.client_id)
id_f_map[mapping].append(fault.fault_address)
id_fx_map[mapping].append(x)
x += 1
bl = bl#[0:20]
cap = bl[-1]
print("plotting vlines")
ylim2 = max(itertools.chain.from_iterable(id_f_map.values()))
ax.vlines(bl, min(ranges), ylim2, label="batch", linewidth=.1)
print("plotting hlines")
#ax.hlines(ranges, 0, len(faults), label="cudaMallocManaged()", linestyles="dotted")
ax.hlines(ranges, 0, cap, label="cudaMallocManaged()", linestyles="dotted")
total_xs = []
for i, (sm_id, color) in enumerate(zip(sm_ids, colors)):
print("plotting sm", sm_id)
#ax.scatter(id_fx_map[sm_id], id_f_map[sm_id] , color=color, label=str(sm_id), s=.01)
x = id_fx_map[sm_id]
y = id_f_map[sm_id]
x = [f for f in x if f < cap]
y = y[0:len(x)]
total_xs += x
print("faults len:", len(x))
#if (sm_id == (46,36)):
#print("x, y for weird SM_ID:", x, ",", y)
#continue
ax.plot(x, y, color=color, label=str(sm_id), linestyle="None", markersize=1, marker="*")
#ax.plot(id_fx_map[sm_id], id_f_map[sm_id] , color=color, label=str(sm_id), linestyle="None", markersize=1)
#if i == 0:
# break
#STARTO = 1606348807401491424
STARTO = 1606352856887328736
ENDO = 1606352856889447392
print ("sorting times lol")
sorted_faults = sorted(faults, key=lambda f: f.timestamp)
print ('building time lists')
newfaults = [f.fault_address for f in sorted_faults]
newx = [f.timestamp for f in sorted_faults]
#newx = [int(f - STARTO) for f in newx]
newx = [int(f - newx[0]) for f in newx]
print("plotting times")
ax2.plot(newx[0: max(total_xs)], newfaults[0:max(total_xs)], 'g'+m, markersize=1, label="time_order")
#ax2.plot([0, ENDO - STARTO], [min(newfaults), min(newfaults)], 'b+', markersize=5, linestyle='None', label="kernel-ret")
temp = newx[0: max(total_xs)]
temp2 = []
for i, f in enumerate(temp):
if i == 0:
temp2.append(f)
else:
if f > temp[i-1] + 1e5:
print ("temporal batch:", len(temp2))
temp2 = [f]
else:
temp2.append(f)
print ("temporal batch:", len(temp2))
for batch in batches:
#for batch in batches[0:20]:
print ("real batchlen:", len(batch))
print ("total batches:", len(batches))
#print([ENDO - STARTO])
#print("newx", newx)
#fig.set_size_inches(16, 10)
ax.set_xlabel('Fault Occurence')
ax.set_ylabel('Fault Fault Index')
psize = dirname(args.csv).split("_")[1]
figname = splitext(basename(args.csv))[0].split("_")[0] + "-" + psize + "-tlb-batch.png"
ylabels = map(lambda t: '0x%013X' % int(t), ax.get_yticks())
ax.set_yticklabels(ylabels)
h1, l1 = ax.get_legend_handles_labels()
h2, l2 = ax2.get_legend_handles_labels()
ax.legend(h1+h2, l1+l2, loc="upper left")
ax2.set_xlabel("Time (NS)");
plt.tight_layout()
print('saving figure:', figname)
fig.savefig(figname, dpi=500)
plt.close(fig)
def main():
parser = argparse.ArgumentParser()
parser.add_argument('csv', type=str, help='Full path to CSV file')
parser.add_argument('-o', type=str, default="", help='output filename')
parser.add_argument('-d', type=str, default="", help='out dir')
args = parser.parse_args()
c = args.csv
m = "*"
if not ".txt" in args.csv:
print("Suspicious input file with no/wrong extension:", args.csv)
exit(1)
e = Experiment(c)
ranges, faults, batches, num_batches = e.get_raw_data()
e.print_info()
matplotlib.rcParams['agg.path.chunksize'] = 10000
#fig = plt.gcf()
#ax = fig.add_subplot(1,1,1)
#ax2 = ax.twiny()
read = []
rx = []
write = []
wx = []
for i, f in enumerate(faults):
if f.access_type == "r":
#read.append(f.fault_address)
#rx.append(i)
pass
elif f.access_type == "w":
#write.append(f.fault_address)
#wx.append(i)
pass
elif f.access_type == "p":
pass
else:
print("unaccounted for access type:", f.access_type)
true_bl = [len(batch) for batch in batches]
bl = [0] + true_bl
for i, b in enumerate(bl):
if i == 0:
continue
bl[i] += bl[i - 1]
del bl[0]
#print(batches)
sm_ids = sorted(e.count_utlb_client_pairs().keys())
print("sm_ids:", sm_ids)
cmap = plt.get_cmap('jet')
colors = cmap(np.linspace(0, 1.0, len(sm_ids)))
id_f_map = {sm_id: [] for sm_id in sm_ids}
for fault in faults:
mapping = (fault.utlb_id, fault.client_id)
id_f_map[mapping].append(fault)
times_between_batches = e.batch_times
avg = e.avg_time_batches()
print("avg time between batches:", avg)
print("min time between batches", min(times_between_batches))
print("max time between batches", max(times_between_batches))
print("total # deltas:", len(times_between_batches))
print("Num batches < 1e1:",
len([t for t in times_between_batches if t < 1e1]))
print("Num batches < 1e2:",
len([t for t in times_between_batches if t < 1e2]))
print("Num batches < 1e3:",
len([t for t in times_between_batches if t < 1e3]))
print("Num batches < 1e4:",
len([t for t in times_between_batches if t < 1e4]))
print("Num batches < 1e5:",
len([t for t in times_between_batches if t < 1e5]))
print("Num batches < 1e6:",
len([t for t in times_between_batches if t < 1e6]))
print("Num batches < 1e7:",
len([t for t in times_between_batches if t < 1e7]))
print("Num batches < 1e8:",
len([t for t in times_between_batches if t < 1e8]))
Q1 = np.quantile(times_between_batches, 0.25)
Q3 = np.quantile(times_between_batches, 0.75)
med = statistics.median(times_between_batches)
avg = np.mean(times_between_batches)
print("Q1, median, Q3:", Q1, ",", med, ",", Q3)
#plt.plot(times, counts, marker="*")
hist, bins, _ = plt.hist(times_between_batches)
#hist, bins, _ = plt.hist(times_between_batches, bins=16)
binlen = len(bins)
print("bins:", bins)
logbins = np.logspace(0.0, np.log10(bins[-1]), len(bins))
print("logbins:", logbins)
plt.clf()
hist, bins, _ = plt.hist(times_between_batches, bins=logbins)
ax = plt.gca()
plt.vlines([Q1, med, Q3],
0,
1,
transform=ax.get_xaxis_transform(),
label="Q1/Med/Q3")
plt.vlines([avg],
0,
1,
transform=ax.get_xaxis_transform(),
label="Avg",
color="r")
plt.xlim(xmin=1e0)
plt.ylim(ymin=0.0)
plt.xscale("log")
plt.xlabel("Time Between Batch Fault Arrival in Buffer (NS)")
plt.ylabel("Frequency")
plt.legend()
psize = basename(dirname(args.csv)) #.split("_")[-1]
print("psize:", psize)
figname = None
if args.o == "":
figname = (args.d + "/" + splitext(basename(args.csv))[0] + "-" +
psize + "-batch-time-dist.png").replace("_", "-")
else:
figname = args.o
if ".png" not in figname:
figname += ".png"
#figname = splitext(basename(args.csv))[0].split("_")[0] + "-" + psize + "-sm-time-dist.png"
plt.tight_layout()
print('saving figure:', figname)
plt.savefig(figname, dpi=500)
plt.close()
if ("pf") in args.csv:
xs = [
len(batch) - dup + len(pfbatch) for batch, dup, pfbatch in zip(
e.batches, e.get_duplicate_faults_64k(), e.pfbatches)
]
else:
xs = [
len(batch) - dup
for batch, dup in zip(e.batches, e.get_duplicate_faults_4k())
]
#xs = [len(batch) - e.get_duplicate_faults_4k() for batch in e.batches]
hist, bins, _ = plt.hist(xs, bins=len(logbins))
logbins = np.logspace(0.0, np.log10(bins[-1]), len(bins))
plt.clf()
hist, bins, _ = plt.hist(xs, bins=logbins)
#plt.hist([len(batch) for batch in e.batches])
plt.xlabel("Batch Sizes")
plt.ylabel("Frequency")
plt.xlim(xmin=1e0)
plt.ylim(ymin=0.0)
plt.xscale("log")
if args.o == "":
figname = (args.d + "/" + splitext(basename(args.csv))[0] + "-" +
psize + "-batch-size-dist.png").replace("_", "-")
else:
figname = args.o
if ".png" not in figname:
figname += ".png"
plt.tight_layout()
print('saving figure:', figname)
plt.savefig(figname, dpi=500)
plt.close()
ys = times_between_batches
size_time_plot(xs, ys, psize, args, "pf-dups")
if ("pf") in args.csv:
xs = [
len(batch) - dup
for batch, dup in zip(e.batches, e.get_duplicate_faults_64k())
]
else:
xs = [
len(batch) - dup
for batch, dup in zip(e.batches, e.get_duplicate_faults_4k())
]
#xs = [len(batch) - e.get_duplicate_faults_4k() for batch in e.batches]
size_time_plot(xs, ys, psize, args, "dups")
if ("pf") in args.csv:
xs = [len(batch) for batch in e.batches]
else:
xs = [len(batch) for batch in e.batches]
size_time_plot(xs, ys, psize, args, "")
div = 65536 if "pf" in args.csv else 4096
xs = []
for batch in batches:
vals = sorted(set([f.fault_address // div for f in batch]))
transfers = 0
prev = vals[0]
count = 0
for val in vals:
if val > 1 + prev or val % (2097152 // div) == 0:
transfers += 1
count = 0
else:
count += 1
prev = val
if count > 0:
transfers += 1
xs.append(transfers)
size_time_plot(xs, ys, psize, args, "transfers")
def main():
parser = argparse.ArgumentParser()
parser.add_argument('csv', type=str, help='Full path to CSV file')
args = parser.parse_args()
c = args.csv
m = "*"
if not ".txt" in args.csv:
print("Suspicious input file with no/wrong extension:", args.csv)
exit(1)
e = Experiment(c)
ranges, faults, batches, num_batches = e.get_raw_data()
e.print_info()
matplotlib.rcParams['agg.path.chunksize'] = 10000
fig = plt.figure()
#fig = plt.gcf()
ax = fig.add_subplot(1, 1, 1)
ax2 = ax.twiny()
read = []
rx = []
write = []
wx = []
for i, f in enumerate(faults):
if f.access_type == "r":
read.append(f.fault_address)
rx.append(i)
elif f.access_type == "w":
write.append(f.fault_address)
wx.append(i)
else:
print("unaccounted for access type:", faults.access_type)
bl = [0] + [len(batch) for batch in batches]
for i, b in enumerate(bl):
if i == 0:
continue
bl[i] += bl[i - 1]
del bl[0]
#print(batches)
print("batch lengths:", bl)
#print(ranges)
ylim1 = min(read, write)
ylim2 = max(max(read), max(write))
ax.vlines(bl, min(ranges), ylim2, label="batch")
ax.hlines(ranges, 0, len(faults), label="cudaMallocManaged()")
ax.plot(rx, read, 'b' + m, markersize=1, label="read")
ax.plot(wx, write, 'r' + m, markersize=1, label="write")
newfaults = [
f.fault_address for f in sorted(faults, key=lambda f: f.timestamp)
]
newx = [f.timestamp for f in sorted(faults, key=lambda f: f.timestamp)]
newx = [int(f - min(newx)) for f in newx]
ax2.plot(newx, newfaults, 'g' + m, markersize=1, label="time_order")
print("newx", newx)
#fig.set_size_inches(16, 10)
ax.set_xlabel('Fault Occurence')
ax.set_ylabel('Fault Fault Index')
psize = dirname(args.csv).split("_")[1]
figname = splitext(basename(
args.csv))[0].split("_")[0] + "-" + psize + "-batch.png"
ylabels = map(lambda t: '0x%013X' % int(t), ax.get_yticks())
ax.set_yticklabels(ylabels)
h1, l1 = ax.get_legend_handles_labels()
h2, l2 = ax2.get_legend_handles_labels()
ax.legend(h1 + h2, l1 + l2, loc="upper left")
ax2.set_xlabel("Time (NS)")
plt.tight_layout()
print('saving figure:', figname)
fig.savefig(figname, dpi=500)
plt.close(fig)