def show_core_locality(df, task_re, label):
df = df[df['event'] == 'sched__sched_switch']
df = df[df['task_name'].str.match(task_re)]
# aggregate all the per core tasks (e.g. swapper/0 -> swapper)
df['task_name'] = df['task_name'].str.replace(r'/.*$', '')
# group by task name
gb = df.groupby('task_name')
task_list = gb.groups
p = figure(plot_width=1000, plot_height=800, **title_style)
p.xaxis.axis_label = 'time (usecs)'
p.yaxis.axis_label = 'core'
p.legend.orientation = "bottom_right"
p.xaxis.axis_label_text_font_size = "10pt"
p.yaxis.axis_label_text_font_size = "10pt"
p.title = "Core locality (%s)" % (label)
color_list = cycle_colors(task_list)
for task, color in zip(task_list, color_list):
dfe = gb.get_group(task)
# add 1 column to contain the starting time for each run period
dfe['start'] = dfe['usecs'] - dfe['duration']
tid = dfe['pid'].iloc[0]
count = len(dfe)
legend_text = '%s:%d (%d)' % (task, tid, count)
# draw end of runs
p.circle('usecs',
'cpu',
source=ColumnDataSource(dfe),
size=get_disc_size(count) + 2,
color=color,
alpha=0.3,
legend=legend_text)
# draw segments to show the entire runs
p.segment('start',
'cpu',
'usecs',
'cpu',
line_width=5,
line_color=color,
source=ColumnDataSource(dfe))
# specify how to output the plot(s)
output_html(p, 'coreloc', task_re)
def show_core_locality(df, task_re, label):
df = df[df['event'] == 'sched__sched_switch']
df = df[df['task_name'].str.match(task_re)]
# aggregate all the per core tasks (e.g. swapper/0 -> swapper)
df['task_name'] = df['task_name'].str.replace(r'/.*$', '')
# group by task name
gb = df.groupby('task_name')
task_list = gb.groups
p = figure(plot_width=1000, plot_height=800, **title_style)
p.xaxis.axis_label = 'time (usecs)'
p.yaxis.axis_label = 'core'
p.legend.orientation = "bottom_right"
p.xaxis.axis_label_text_font_size = "10pt"
p.yaxis.axis_label_text_font_size = "10pt"
p.title = "Core locality (%s)" % (label)
color_list = cycle_colors(task_list)
for task, color in zip(task_list, color_list):
dfe = gb.get_group(task)
# add 1 column to contain the starting time for each run period
dfe['start'] = dfe['usecs'] - dfe['duration']
tid = dfe['pid'].iloc[0]
count = len(dfe)
legend_text = '%s:%d (%d)' % (task, tid, count)
# draw end of runs
p.circle('usecs', 'cpu', source=ColumnDataSource(dfe),
size=get_disc_size(count) + 2, color=color,
alpha=0.3,
legend=legend_text)
# draw segments to show the entire runs
p.segment('start', 'cpu', 'usecs', 'cpu', line_width=5, line_color=color,
source=ColumnDataSource(dfe))
# specify how to output the plot(s)
output_html('coreloc', task_re)
# display the figure
show(p)
def show_sw_kvm_heatmap(df, task_re, label, show_ctx_switches, show_kvm):
gb = get_groupby(df, task_re)
chart_list = []
# these are the 2 main events to show for kvm events
legend_map_kvm = {
'kvm_exit': (BLUE, 'vcpu running (y=vcpu run time)', False),
'kvm_entry': (ORANGE, 'vcpu not running (y=kvm+sleep time)', False)
}
# context switch events
legend_map_ctx_sw = {
'sched__sched_stat_sleep': (RED, 'wakeup from sleep (y=sleep time)', True),
'sched__sched_switch': (GREEN, 'switched out from cpu (y=run time)', True)
}
if show_kvm and show_ctx_switches:
legend_map = dict(legend_map_kvm.items() + legend_map_ctx_sw.items())
title = "Scheduler and KVM events"
prefix = 'swkvm'
elif show_kvm:
legend_map = legend_map_kvm
title = "KVM events"
prefix = 'kvm'
else:
legend_map = legend_map_ctx_sw
title = "Scheduler events"
prefix = 'sw'
width = 1000
height = 800
show_legend = True
nb_charts = len(gb.groups)
if nb_charts == 0:
print 'No selection matching: ' + task_re
return
if nb_charts > 1:
width /= 2
height /= 2
tstyle = grid_title_style
else:
tstyle = title_style
task_list = gb.groups.keys()
task_list.sort()
show_legend = True
duration_max = usecs_max = -1
duration_min = usecs_min = sys.maxint
event_list = legend_map.keys()
event_list.sort()
for task in task_list:
p = figure(plot_width=width, plot_height=height, y_axis_type="log", **tstyle)
p.xaxis.axis_label = 'time (usecs)'
p.yaxis.axis_label = 'duration (usecs)'
p.legend.orientation = "bottom_right"
p.xaxis.axis_label_text_font_size = "10pt"
p.yaxis.axis_label_text_font_size = "10pt"
if label:
p.title = "%s for %s (%s)" % (title, task, label)
label = None
else:
p.title = task
p.ygrid.minor_grid_line_color = 'navy'
p.ygrid.minor_grid_line_alpha = 0.1
accumulated_time = {}
total_time = 0
dfg = gb.get_group(task)
# remove any row with zero duration as it confuses the chart library
dfg = dfg[dfg['duration'] > 0]
for event in event_list:
dfe = dfg[dfg.event == event]
duration_min = min(duration_min, dfe['duration'].min())
duration_max = max(duration_max, dfe['duration'].max())
usecs_min = min(usecs_min, dfe['usecs'].min())
usecs_max = max(usecs_max, dfe['usecs'].max())
count = len(dfe)
color, legend_text, cx_sw = legend_map[event]
if show_legend:
legend_text = '%s (%d)' % (legend_text, count)
elif color == GREEN:
legend_text = '(%d)' % (count)
else:
legend_text = None
# there is bug in bokeh when there are too many circles to draw, nothing is visible
if len(dfe) > 50000:
dfe = dfe[:50000]
print 'Series for %s display truncated to 50000 events' % (event)
if cx_sw:
draw_shape = p.circle
size = get_disc_size(count)
else:
draw_shape = p.diamond
size = get_disc_size(count) + 4
draw_shape('usecs', 'duration', source=ColumnDataSource(dfe),
size=size,
color=color,
alpha=0.3,
legend=legend_text)
event_duration = dfe['duration'].sum()
accumulated_time[event] = event_duration
total_time += event_duration
chart_list.append(p)
show_legend = False
shared_x_range = Range1d(usecs_min, usecs_max)
shared_y_range = Range1d(duration_min, duration_max)
for p in chart_list:
p.x_range = shared_x_range
p.y_range = shared_y_range
# specify how to output the plot(s)
output_html(prefix, task_re)
# display the figure
if len(chart_list) == 1:
show(chart_list[0])
else:
# split the list into an array of rows with 2 charts per row
gp = gridplot(split_list(chart_list, 2))
show(gp)
def show_core_runs(df, task_re, label, duration):
time_span_msec = get_time_span_msec(df)
# remove unneeded columns
df = df.drop('next_pid', axis=1)
df.drop('pid', axis=1, inplace=True)
df.drop('usecs', axis=1, inplace=True)
df.drop('next_comm', axis=1, inplace=True)
# filter out all events except the switch events
df = df[df.event == 'sched__sched_switch']
df = df.drop('event', axis=1)
df = df[df['task_name'].str.match(task_re)]
# at this point we have a df that looks like this:
# task_name cpu duration
# 0 ASA.1.vcpu0 8 7954
# 1 ASA.1.vcpu0 9 5475
# 2 ASA.1.vcpu0 10 4151
# 3 ASA.1.vcpu0 11 12391
# 4 ASA.1.vcpu0 12 21025
# 5 ASA.1.vcpu0 13 6447
# 6 ASA.1.vcpu0 14 16798
# 7 ASA.1.vcpu0 15 3911
# 8 ASA.10.vcpu0 8 4248
# 9 ASA.10.vcpu0 9 3534
# 10 ASA.10.vcpu0 10 15624
# 11 ASA.10.vcpu0 11 6925
# etc...
if len(df) == 0:
print
print 'No selection matching "%s"' % (task_re)
return
gb = df.groupby(['task_name', 'cpu'], as_index=False)
if duration:
# add duration values
df = gb.aggregate(np.sum)
max_core = df.cpu.max()
dfsum = df.drop('cpu', axis=1)
gb = dfsum.groupby('task_name', as_index=False)
dfsum = gb.aggregate(np.sum)
# dfsum is the sum of all duration for given task
# 0 ASA.1.vcpu0 78152
# 1 ASA.10.vcpu0 65637
# 2 ASA.11.vcpu0 81525
# 3 ASA.12.vcpu0 56488
# For each task, the maximum runtime is the time_span_msec (100% of 1 core)
# The idle time for each task is therefore time_span_msec * 1000 - sum(duration)
dfsum['cpu'] = 'IDLE'
time_span_usec = time_span_msec * 1000
dfsum['duration'] = time_span_usec - dfsum['duration']
# now we need to reinsert that data back to the df
dfm = pandas.concat([df, dfsum], ignore_index=True)
# task_name cpu duration total
# 0 ASA.1.vcpu0 8 7954 78152
# 1 ASA.1.vcpu0 9 5475 78152
# 2 ASA.1.vcpu0 10 4151 78152
# Add a % column
dfm['percent'] = ((dfm['duration'] * 100 * 10) // time_span_usec) / 10
# This is for the legend
min_count = 0
max_count = 100
range_unit = '%'
# many core-pinned system tasks have a duration of 0 (swapper, watchdog...)
dfm.fillna(100, inplace=True)
dfm.drop(['duration'], axis=1, inplace=True)
tooltip_count = ("time", "@percent% of core @cpu")
title = "Task Run Time %% per Core (%s, %d msec window)" % (label, time_span_msec)
# this is YlGnBu9[::-1] (Reverse the color order so dark is highest value)
# with an extra intermediate color to make it 10
palette = ['#ffffd9', '#edf8b1', '#c7e9b4', '#a3dbb7', '#7fcdbb',
'#41b6c4', '#1d91c0', '#225ea8', '#253494', '#081d58']
html_prefix = 'core_runtime'
# add 1 extra column in the heatmap for core "IDLE" to represent the IDLE time
# this requires adding 1 row per real core that has a runtime set to
# capture window size - sum(duration on all cores)
else:
# count number of rows with same task and cpu
dfm = DataFrame(gb.size())
dfm.reset_index(inplace=True)
dfm.rename(columns={0: 'count'}, inplace=True)
min_count = dfm['count'].min()
max_count = dfm['count'].max()
range_unit = ''
spread = max_count - min_count
# Add a % column
dfm['percent'] = ((dfm['count'] - min_count) * 100) / spread
tooltip_count = ("context switches", "@count")
title = "Task Context Switches per Core (%s, %d msec window)" % (label, time_span_msec)
palette = YlOrRd9[::-1]
html_prefix = 'core_switches'
max_core = dfm.cpu.max()
# get the list of cores
# round up to next mutliple of 4 - 1
# 0..3 -> 3
# 4..7 -> 7 etc
max_core |= 0x03
max_core = max(max_core, 3)
core_list = [str(x) for x in range(max_core + 1)]
if duration:
core_list.append('IDLE')
# make room for the legend by adding 3 empty columns
core_list += ['', '', '']
dfm['cpu'] = dfm['cpu'].astype(str)
# replace ':' with '_' as it would cause bokeh to misplace the labels on the chart
dfm['task_name'] = dfm['task_name'].str.replace(':', '_')
normalize_df_task_name(dfm)
# Add a column for the Y axis
# each task name should be associated to a unique Y index
# Get a unique list of task names sorted
task_list = pandas.unique(dfm.task_name.ravel()).tolist()
task_list.sort()
# Add a color column
dfm['color'] = dfm.apply(lambda row: get_color(row['percent'],
palette), axis=1)
# switch to str type to prevent the tooltip to
# display percent value with 3 digits
dfm.percent = dfm.percent.astype(str)
# make enough vertical space for the legend
# the legend needs 1 row per palette color + 1 to fit the max value
# so we need at least len(palette) + 1 rows
if len(task_list) < len(palette) + 1:
task_list += ['' for _ in range(len(palette) + 1 - len(task_list))]
TOOLS = "resize,hover,save"
p = figure(title=title, tools=TOOLS, x_range=core_list, y_range=task_list, **title_style)
p.plot_width = 1000
p.plot_height = 80 + len(task_list) * 16
p.toolbar_location = "left"
source = ColumnDataSource(dfm)
# the name is to flag these rectangles to enable tooltip hover on them
# (and not enable tooltips on the legend patches)
p.rect("cpu", "task_name", width=1, height=0.9, source=source, fill_alpha=0.8, color="color", name='patches')
p.grid.grid_line_color = None
# trace separator lines to isolate blocks across core groups (numa sockets) and task-like names
max_y = len(task_list)
# trace a vertical line every 8 cores
for seg_x in range(8, max_core + 7, 8):
p.segment(x0=[seg_x + 0.5], y0=[0], x1=[seg_x + 0.5],
y1=[max_y + 0.5], color=GRAY, line_width=2)
prev_task_name = None
# trace a horizontal line around every group of tasks that have the same first 3 characters
for y in range(max_y):
cur_task_name = task_list[y]
if prev_task_name:
if prev_task_name[0:3] != cur_task_name[0:3]:
p.segment(x0=[0], y0=[y + 0.5], x1=[max_core + 1.5],
y1=[y + 0.5], color=GRAY, line_width=0.5)
prev_task_name = task_list[y]
hover = p.select(dict(type=HoverTool))
hover.tooltips = OrderedDict([
("task", "@task_name"),
("core", "@cpu"),
tooltip_count
])
# only enable tooltip on rectangles with name 'patches'
hover.names = ['patches']
# legend to the right
# we try to center the legend vertically
legend_base_y = 0
palette_len = len(palette)
if max_y > palette_len + 1:
legend_base_y = (max_y - palette_len - 1) // 2
if duration:
# IDLE cpu inserted so shift the legend by 1 position to the right
max_core += 1
# pass 1 is to draw the color patches
# prepare a data source with a x, y and a color column
x_values = np.empty(palette_len)
x_values.fill(max_core + 2.5)
y_values = np.arange(legend_base_y + 1.5, legend_base_y + palette_len + 1, 1)
dfl = DataFrame({'x': x_values, 'y': y_values, 'color': palette})
source = ColumnDataSource(dfl)
p.rect(x='x', y='y', color='color', width=1, height=1, source=source)
# pass 2 is to draw the text describing the ranges for the color patches
color_value_list = get_color_value_list(min_count, max_count, palette, range_unit)
x_values = np.empty(len(color_value_list))
x_values.fill(max_core + 3.1)
y_values = np.arange(legend_base_y + 0.7, legend_base_y + len(color_value_list), 1)
dfl = DataFrame({'x': x_values, 'y': y_values, 'color_values': color_value_list})
source = ColumnDataSource(dfl)
p.text(x='x', y='y', text='color_values', source=source,
text_font_size='8pt')
output_html(p, html_prefix, task_re)
def show_core_runs(df, task_re, label, duration):
time_span_msec = get_time_span_msec(df)
# remove unneeded columns
df = df.drop('next_pid', axis=1)
df.drop('pid', axis=1, inplace=True)
df.drop('usecs', axis=1, inplace=True)
df.drop('next_comm', axis=1, inplace=True)
# filter out all events except the switch events
df = df[df.event == 'sched__sched_switch']
df = df.drop('event', axis=1)
df = df[df['task_name'].str.match(task_re)]
# at this point we have a df that looks like this:
# task_name cpu duration
# 0 ASA.1.vcpu0 8 7954
# 1 ASA.1.vcpu0 9 5475
# 2 ASA.1.vcpu0 10 4151
# 3 ASA.1.vcpu0 11 12391
# 4 ASA.1.vcpu0 12 21025
# 5 ASA.1.vcpu0 13 6447
# 6 ASA.1.vcpu0 14 16798
# 7 ASA.1.vcpu0 15 3911
# 8 ASA.10.vcpu0 8 4248
# 9 ASA.10.vcpu0 9 3534
# 10 ASA.10.vcpu0 10 15624
# 11 ASA.10.vcpu0 11 6925
# etc...
if len(df) == 0:
print
print 'No selection matching "%s"' % (task_re)
return
gb = df.groupby(['task_name', 'cpu'], as_index=False)
if duration:
# add duration values
df = gb.aggregate(np.sum)
max_core = df.cpu.max()
dfsum = df.drop('cpu', axis=1)
gb = dfsum.groupby('task_name', as_index=False)
dfsum = gb.aggregate(np.sum)
# dfsum is the sum of all duration for given task
# 0 ASA.1.vcpu0 78152
# 1 ASA.10.vcpu0 65637
# 2 ASA.11.vcpu0 81525
# 3 ASA.12.vcpu0 56488
# For each task, the maximum runtime is the time_span_msec (100% of 1 core)
# The idle time for each task is therefore time_span_msec * 1000 - sum(duration)
dfsum['cpu'] = 'IDLE'
time_span_usec = time_span_msec * 1000
dfsum['duration'] = time_span_usec - dfsum['duration']
# now we need to reinsert that data back to the df
dfm = pandas.concat([df, dfsum], ignore_index=True)
# task_name cpu duration total
# 0 ASA.1.vcpu0 8 7954 78152
# 1 ASA.1.vcpu0 9 5475 78152
# 2 ASA.1.vcpu0 10 4151 78152
# Add a % column
dfm['percent'] = ((dfm['duration'] * 100 * 10) // time_span_usec) / 10
# This is for the legend
min_count = 0
max_count = 100
range_unit = '%'
# many core-pinned system tasks have a duration of 0 (swapper, watchdog...)
dfm.fillna(100, inplace=True)
dfm.drop(['duration'], axis=1, inplace=True)
tooltip_count = ("time", "@percent% of core @cpu")
title = "Task Run Time %% per Core (%s, %d msec window)" % (
label, time_span_msec)
# this is YlGnBu9[::-1] (Reverse the color order so dark is highest value)
# with an extra intermediate color to make it 10
palette = [
'#ffffd9', '#edf8b1', '#c7e9b4', '#a3dbb7', '#7fcdbb', '#41b6c4',
'#1d91c0', '#225ea8', '#253494', '#081d58'
]
html_prefix = 'core_runtime'
# add 1 extra column in the heatmap for core "IDLE" to represent the IDLE time
# this requires adding 1 row per real core that has a runtime set to
# capture window size - sum(duration on all cores)
else:
# count number of rows with same task and cpu
dfm = DataFrame(gb.size())
dfm.reset_index(inplace=True)
dfm.rename(columns={0: 'count'}, inplace=True)
min_count = dfm['count'].min()
max_count = dfm['count'].max()
range_unit = ''
spread = max_count - min_count
# Add a % column
dfm['percent'] = ((dfm['count'] - min_count) * 100) / spread
tooltip_count = ("context switches", "@count")
title = "Task Context Switches per Core (%s, %d msec window)" % (
label, time_span_msec)
palette = YlOrRd9[::-1]
html_prefix = 'core_switches'
max_core = dfm.cpu.max()
# get the list of cores
# round up to next mutliple of 4 - 1
# 0..3 -> 3
# 4..7 -> 7 etc
max_core |= 0x03
max_core = max(max_core, 3)
core_list = [str(x) for x in range(max_core + 1)]
if duration:
core_list.append('IDLE')
# make room for the legend by adding 3 empty columns
core_list += ['', '', '']
dfm['cpu'] = dfm['cpu'].astype(str)
# replace ':' with '_' as it would cause bokeh to misplace the labels on the chart
dfm['task_name'] = dfm['task_name'].str.replace(':', '_')
normalize_df_task_name(dfm)
# Add a column for the Y axis
# each task name should be associated to a unique Y index
# Get a unique list of task names sorted
task_list = pandas.unique(dfm.task_name.ravel()).tolist()
task_list.sort()
# Add a color column
dfm['color'] = dfm.apply(lambda row: get_color(row['percent'], palette),
axis=1)
# switch to str type to prevent the tooltip to
# display percent value with 3 digits
dfm.percent = dfm.percent.astype(str)
# make enough vertical space for the legend
# the legend needs 1 row per palette color + 1 to fit the max value
# so we need at least len(palette) + 1 rows
if len(task_list) < len(palette) + 1:
task_list += ['' for _ in range(len(palette) + 1 - len(task_list))]
TOOLS = "resize,hover,save"
p = figure(title=title,
tools=TOOLS,
x_range=core_list,
y_range=task_list,
**title_style)
p.plot_width = 1000
p.plot_height = 80 + len(task_list) * 16
p.toolbar_location = "left"
source = ColumnDataSource(dfm)
# the name is to flag these rectangles to enable tooltip hover on them
# (and not enable tooltips on the legend patches)
p.rect("cpu",
"task_name",
width=1,
height=0.9,
source=source,
fill_alpha=0.8,
color="color",
name='patches')
p.grid.grid_line_color = None
# trace separator lines to isolate blocks across core groups (numa sockets) and task-like names
max_y = len(task_list)
# trace a vertical line every 8 cores
for seg_x in range(8, max_core + 7, 8):
p.segment(x0=[seg_x + 0.5],
y0=[0],
x1=[seg_x + 0.5],
y1=[max_y + 0.5],
color=GRAY,
line_width=2)
prev_task_name = None
# trace a horizontal line around every group of tasks that have the same first 3 characters
for y in range(max_y):
cur_task_name = task_list[y]
if prev_task_name:
if prev_task_name[0:3] != cur_task_name[0:3]:
p.segment(x0=[0],
y0=[y + 0.5],
x1=[max_core + 1.5],
y1=[y + 0.5],
color=GRAY,
line_width=0.5)
prev_task_name = task_list[y]
hover = p.select(dict(type=HoverTool))
hover.tooltips = OrderedDict([("task", "@task_name"), ("core", "@cpu"),
tooltip_count])
# only enable tooltip on rectangles with name 'patches'
hover.names = ['patches']
# legend to the right
# we try to center the legend vertically
legend_base_y = 0
palette_len = len(palette)
if max_y > palette_len + 1:
legend_base_y = (max_y - palette_len - 1) // 2
if duration:
# IDLE cpu inserted so shift the legend by 1 position to the right
max_core += 1
# pass 1 is to draw the color patches
# prepare a data source with a x, y and a color column
x_values = np.empty(palette_len)
x_values.fill(max_core + 2.5)
y_values = np.arange(legend_base_y + 1.5, legend_base_y + palette_len + 1,
1)
dfl = DataFrame({'x': x_values, 'y': y_values, 'color': palette})
source = ColumnDataSource(dfl)
p.rect(x='x', y='y', color='color', width=1, height=1, source=source)
# pass 2 is to draw the text describing the ranges for the color patches
color_value_list = get_color_value_list(min_count, max_count, palette,
range_unit)
x_values = np.empty(len(color_value_list))
x_values.fill(max_core + 3.1)
y_values = np.arange(legend_base_y + 0.7,
legend_base_y + len(color_value_list), 1)
dfl = DataFrame({
'x': x_values,
'y': y_values,
'color_values': color_value_list
})
source = ColumnDataSource(dfl)
p.text(x='x',
y='y',
text='color_values',
source=source,
text_font_size='8pt')
output_html(p, html_prefix, task_re)
def show_kvm_exit_types(df, task_re, label):
df = df[df['event'] == 'kvm_exit']
df = df[df['task_name'].str.match(task_re)]
# the next_comm column contains the exit code
exit_codes = Series(KVM_EXIT_REASONS)
# add new column congaining the exit reason in clear text
df['exit_reason'] = df['next_comm'].map(exit_codes)
time_span_msec = get_time_span_msec(df)
df.drop(['cpu', 'duration', 'event', 'next_pid', 'pid', 'next_comm', 'usecs'], inplace=True, axis=1)
# Get the list of exit reasons, sorted alphabetically
reasons = pandas.unique(df.exit_reason.ravel()).tolist()
reasons.sort()
# group by task name then exit reasons
gb = df.groupby(['task_name', 'exit_reason'])
# number of exit types
size_series = gb.size()
df = size_series.to_frame('count')
df.reset_index(inplace=True)
p = Bar(df, label='task_name', values='count', stack='exit_reason',
title="KVM Exit types per task (%s, %d msec window)" % (label, time_span_msec),
legend='top_right',
tools="resize,hover,save",
width=1000, height=800)
p._xaxis.axis_label = "Task Name"
p._xaxis.axis_label_text_font_size = "12pt"
p._yaxis.axis_label = "Exit Count (sum)"
p._yaxis.axis_label_text_font_size = "12pt"
# Cannot find a way to display the exit reason in the tooltip
# from bokeh.models.renderers import GlyphRenderer
# glr = p.select(dict(type=GlyphRenderer))
# bar_source = glr[0].data_source
# print bar_source.data
# bar_source = glr[1].data_source
# bar_source.data['exit_reason'] = ['HOHO']
hover = p.select(dict(type=HoverTool))
hover.tooltips = OrderedDict([
("task", "$x"),
# {"reason", "@exit_reason"},
("count", "@height")
])
# specify how to output the plot(s)
# table with counts
gb = df.groupby(['exit_reason'])
keys = gb.groups.keys()
dfr_list = []
for reason in keys:
dfr = gb.get_group(reason)
# drop the exit reason column
dfr = dfr.drop(['exit_reason'], axis=1)
# rename the count column with the reason name
dfr.rename(columns={'count': reason}, inplace=True)
# set the task name as the index
dfr.set_index('task_name', inplace=True)
dfr_list.append(dfr)
# concatenate all task columns into 1 dataframe that has the exit reason as the index
# counts for missing exit reasons will be set to NaN
dft = pandas.concat(dfr_list, axis=1)
dft.fillna(0, inplace=True)
# Add a total column
dft['TOTAL'] = dft.sum(axis=1)
sfmt = StringFormatter(text_align='center', font_style='bold')
nfmt = NumberFormatter(format='0,0')
col_names = list(dft.columns.values)
col_names.sort()
# move 'TOTAL' at end of list
col_names.remove('TOTAL')
col_names.append('TOTAL')
# convert index to column name
dft.reset_index(level=0, inplace=True)
dft.rename(columns={'index': 'Task'}, inplace=True)
columns = [TableColumn(field=name, title=name, formatter=nfmt) for name in col_names]
columns.insert(0, TableColumn(field='Task', title='Task', formatter=sfmt))
table = DataTable(source=ColumnDataSource(dft), columns=columns, width=1000,
row_headers=False,
height='auto')
output_html(vplot(p, table), 'kvm-types', task_re)
'''
def show_sw_kvm_heatmap(df, task_re, label, show_ctx_switches, show_kvm):
gb = get_groupby(df, task_re)
chart_list = []
# these are the 2 main events to show for kvm events
legend_map_kvm = {
'kvm_exit': (BLUE, 'vcpu running (y=vcpu run time)', False),
'kvm_entry': (ORANGE, 'vcpu not running (y=kvm+sleep time)', False)
}
# context switch events
legend_map_ctx_sw = {
'sched__sched_stat_sleep': (RED, 'wakeup from sleep (y=sleep time)', True),
'sched__sched_switch': (GREEN, 'switched out from cpu (y=run time)', True)
}
if show_kvm and show_ctx_switches:
legend_map = dict(legend_map_kvm.items() + legend_map_ctx_sw.items())
title = "Scheduler and KVM events"
prefix = 'swkvm'
elif show_kvm:
legend_map = legend_map_kvm
title = "KVM events"
prefix = 'kvm'
else:
legend_map = legend_map_ctx_sw
title = "Scheduler events"
prefix = 'sw'
width = 1000
height = 800
show_legend = True
nb_charts = len(gb.groups)
if nb_charts == 0:
print 'No selection matching: ' + task_re
return
if nb_charts > 1:
width /= 2
height /= 2
tstyle = grid_title_style
else:
tstyle = title_style
task_list = gb.groups.keys()
task_list.sort()
show_legend = True
duration_max = usecs_max = -1
duration_min = usecs_min = sys.maxint
event_list = legend_map.keys()
event_list.sort()
for task in task_list:
p = figure(plot_width=width, plot_height=height, y_axis_type="log", **tstyle)
p.xaxis.axis_label = 'time (usecs)'
p.yaxis.axis_label = 'duration (usecs)'
p.legend.orientation = "bottom_right"
p.xaxis.axis_label_text_font_size = "10pt"
p.yaxis.axis_label_text_font_size = "10pt"
if label:
p.title = "%s for %s (%s)" % (title, task, label)
label = None
else:
p.title = task
p.ygrid.minor_grid_line_color = 'navy'
p.ygrid.minor_grid_line_alpha = 0.1
accumulated_time = {}
total_time = 0
dfg = gb.get_group(task)
# remove any row with zero duration as it confuses the chart library
dfg = dfg[dfg['duration'] > 0]
for event in event_list:
dfe = dfg[dfg.event == event]
duration_min = min(duration_min, dfe['duration'].min())
duration_max = max(duration_max, dfe['duration'].max())
usecs_min = min(usecs_min, dfe['usecs'].min())
usecs_max = max(usecs_max, dfe['usecs'].max())
count = len(dfe)
color, legend_text, cx_sw = legend_map[event]
if show_legend:
legend_text = '%s (%d)' % (legend_text, count)
elif color == GREEN:
legend_text = '(%d)' % (count)
else:
legend_text = None
# there is bug in bokeh when there are too many circles to draw, nothing is visible
if len(dfe) > 50000:
dfe = dfe[:50000]
print 'Series for %s display truncated to 50000 events' % (event)
if cx_sw:
draw_shape = p.circle
size = get_disc_size(count)
else:
draw_shape = p.diamond
size = get_disc_size(count) + 4
draw_shape('usecs', 'duration', source=ColumnDataSource(dfe),
size=size,
color=color,
alpha=0.3,
legend=legend_text)
event_duration = dfe['duration'].sum()
accumulated_time[event] = event_duration
total_time += event_duration
chart_list.append(p)
show_legend = False
shared_x_range = Range1d(usecs_min, usecs_max)
shared_y_range = Range1d(duration_min, duration_max)
for p in chart_list:
p.x_range = shared_x_range
p.y_range = shared_y_range
# specify how to output the plot(s)
# display the figure
if len(chart_list) == 1:
output_html(chart_list[0], prefix, task_re)
else:
# split the list into an array of rows with 2 charts per row
gp = gridplot(split_list(chart_list, 2))
output_html(gp, prefix, task_re)
def show_kvm_exit_types(df, task_re, label):
df = df[df['event'] == 'kvm_exit']
df = df[df['task_name'].str.match(task_re)]
# the next_comm column contains the exit code
exit_codes = Series(KVM_EXIT_REASONS)
# add new column congaining the exit reason in clear text
df['exit_reason'] = df['next_comm'].map(exit_codes)
time_span_msec = get_time_span_msec(df)
df.drop(
['cpu', 'duration', 'event', 'next_pid', 'pid', 'next_comm', 'usecs'],
inplace=True,
axis=1)
# Get the list of exit reasons, sorted alphabetically
reasons = pandas.unique(df.exit_reason.ravel()).tolist()
reasons.sort()
# group by task name then exit reasons
gb = df.groupby(['task_name', 'exit_reason'])
# number of exit types
size_series = gb.size()
df = size_series.to_frame('count')
df.reset_index(inplace=True)
p = Bar(df,
label='task_name',
values='count',
stack='exit_reason',
title="KVM Exit types per task (%s, %d msec window)" %
(label, time_span_msec),
legend='top_right',
tools="resize,hover,save",
width=1000,
height=800)
p._xaxis.axis_label = "Task Name"
p._xaxis.axis_label_text_font_size = "12pt"
p._yaxis.axis_label = "Exit Count (sum)"
p._yaxis.axis_label_text_font_size = "12pt"
# Cannot find a way to display the exit reason in the tooltip
# from bokeh.models.renderers import GlyphRenderer
# glr = p.select(dict(type=GlyphRenderer))
# bar_source = glr[0].data_source
# print bar_source.data
# bar_source = glr[1].data_source
# bar_source.data['exit_reason'] = ['HOHO']
hover = p.select(dict(type=HoverTool))
hover.tooltips = OrderedDict([
("task", "$x"),
# {"reason", "@exit_reason"},
("count", "@height")
])
# specify how to output the plot(s)
# table with counts
gb = df.groupby(['exit_reason'])
keys = gb.groups.keys()
dfr_list = []
for reason in keys:
dfr = gb.get_group(reason)
# drop the exit reason column
dfr = dfr.drop(['exit_reason'], axis=1)
# rename the count column with the reason name
dfr.rename(columns={'count': reason}, inplace=True)
# set the task name as the index
dfr.set_index('task_name', inplace=True)
dfr_list.append(dfr)
# concatenate all task columns into 1 dataframe that has the exit reason as the index
# counts for missing exit reasons will be set to NaN
dft = pandas.concat(dfr_list, axis=1)
dft.fillna(0, inplace=True)
# Add a total column
dft['TOTAL'] = dft.sum(axis=1)
sfmt = StringFormatter(text_align='center', font_style='bold')
nfmt = NumberFormatter(format='0,0')
col_names = list(dft.columns.values)
col_names.sort()
# move 'TOTAL' at end of list
col_names.remove('TOTAL')
col_names.append('TOTAL')
# convert index to column name
dft.reset_index(level=0, inplace=True)
dft.rename(columns={'index': 'Task'}, inplace=True)
columns = [
TableColumn(field=name, title=name, formatter=nfmt)
for name in col_names
]
columns.insert(0, TableColumn(field='Task', title='Task', formatter=sfmt))
table = DataTable(source=ColumnDataSource(dft),
columns=columns,
width=1000,
row_headers=False,
height='auto')
output_html(vplot(p, table), 'kvm-types', task_re)
'''