Build profiling, now in pictures
We have just merged Beth's initial buildstats logging work. I was sitting wondering how to actually evaluate the numbers as I wanted to know "where are we spending the time?".
It occurred to me that I wanted a graph very similar to that generated by bootchart. I looked around and found pyboootchartgui and then hacked it around a bit and coerced it to start producing charts like:
http://tim.rpsys.net/bootchart.png
which is the initial "pseudo-native" part of the build. This was simple enough to test with.
I then tried graphing a poky-image-sato. To get a graph I could actually read, I stripped out any task taking less than 8 seconds and scaled the x axis from 25 units per second to one unit per second. The result was:
http://tim.rpsys.net/bootchart2.png
(warning this is a 2.7MB png)
I also added in a little bit of colour coding for the second chart. Interestingly it looks like there is more yellow than green meaning configure is a bigger drain on the build time not that its unexpected :/.
I quite enjoyed playing with this and on a serious note, the gradient of the task graph makes me a little suspicious of whether the overhead of launching tasks in bitbake itself is having some effect on build time. Certainly on the first graph there are some interesting latencies showing up.
Anyhow, I think this is the first time bitbake's task execution has been visualised and there are some interesting things we can learn from it. I'm hoping this is a start of a much more detailed understanding of the build process with respect to performance.
Cheers,
Richard
draw.py +91 -39
@@ -53,7 +53,7 @@ # Process border color. PROC_BORDER_COLOR = (0.71, 0.71, 0.71, 1.0) -# Waiting process color. + PROC_COLOR_D = (0.76, 0.48, 0.48, 0.125) # Running process color. PROC_COLOR_R = CPU_COLOR
@@ -88,6 +88,17 @@ # Process description date format. DESC_TIME_FORMAT = "mm:ss.SSS" +# Configure task color +TASK_COLOR_CONFIGURE = (1.0, 1.0, 0.00, 1.0) +# Compile task color. +TASK_COLOR_COMPILE = (0.0, 1.00, 0.00, 1.0) +# Install task color +TASK_COLOR_INSTALL = (1.0, 0.00, 1.00, 1.0) +# Package task color +TASK_COLOR_PACKAGE = (0.0, 1.00, 1.00, 1.0) +# Sysroot task color +TASK_COLOR_SYSROOT = (0.0, 0.00, 1.00, 1.0) + # Process states STATE_UNDEFINED = 0 STATE_RUNNING = 1
@@ -147,7 +158,7 @@ def draw_5sec_labels(ctx, rect, sec_w): ctx.set_font_size(AXIS_FONT_SIZE) for i in range(0, rect[2] + 1, sec_w): - if ((i / sec_w) % 5 == 0) : + if ((i / sec_w) % 30 == 0) : label = "%ds" % (i / sec_w) label_w = ctx.text_extents(label)[2] draw_text(ctx, label, TEXT_COLOR, rect[0] + i - label_w/2, rect[1] - 2)
@@ -159,7 +170,7 @@ ctx.set_line_cap(cairo.LINE_CAP_SQUARE) for i in range(sec_w, rect[2] + 1, sec_w): - if ((i / sec_w) % 5 == 0) : + if ((i / sec_w) % 30 == 0) : ctx.set_source_rgba(*TICK_COLOR_BOLD) else : ctx.set_source_rgba(*TICK_COLOR)
@@ -204,23 +215,26 @@ bar_h = 55 # offsets off_x, off_y = 10, 10 -sec_w = 25 # the width of a second +sec_w = 1 # the width of a second proc_h = 16 # the height of a process leg_s = 10 MIN_IMG_W = 800 OPTIONS = None -def extents(headers, cpu_stats, disk_stats, proc_tree): - w = (proc_tree.duration * sec_w / 100) + 2*off_x - h = proc_h * proc_tree.num_proc + header_h + 2*off_y +def extents(res): + start = min(res.start.keys()) + end = max(res.end.keys()) + + w = ((end - start) * sec_w) + 2*off_x + h = proc_h * len(res.processes) + header_h + 2*off_y return (w,h) # # Render the chart. # -def render(ctx, options, headers, cpu_stats, disk_stats, proc_tree): - (w, h) = extents(headers, cpu_stats, disk_stats, proc_tree) +def render(ctx, options, res): + (w, h) = extents(res) global OPTIONS OPTIONS = options
@@ -231,65 +245,103 @@ draw_fill_rect(ctx, WHITE, (0, 0, max(w, MIN_IMG_W), h)) w -= 2*off_x # draw the title and headers - curr_y = draw_header(ctx, headers, off_x, proc_tree.duration) + #curr_y = draw_header(ctx, headers, off_x, proc_tree.duration) + curr_y = 0 # render bar legend ctx.set_font_size(LEGEND_FONT_SIZE) - draw_legend_box(ctx, "CPU (user+sys)", CPU_COLOR, off_x, curr_y+20, leg_s) - draw_legend_box(ctx, "I/O (wait)", IO_COLOR, off_x + 120, curr_y+20, leg_s) + #print "w, h %s %s" % (w, h) + #draw_legend_box(ctx, "CPU (user+sys)", CPU_COLOR, off_x, curr_y+20, leg_s) + #draw_legend_box(ctx, "I/O (wait)", IO_COLOR, off_x + 120, curr_y+20, leg_s) + # render I/O wait - chart_rect = (off_x, curr_y+30, w, bar_h) - draw_box_ticks(ctx, chart_rect, sec_w) - draw_chart(ctx, IO_COLOR, True, chart_rect, [(sample.time, sample.user + sample.sys + sample.io) for sample in cpu_stats], proc_tree) + #chart_rect = (off_x, curr_y+30, w, bar_h) + #draw_box_ticks(ctx, chart_rect, sec_w) + #draw_chart(ctx, IO_COLOR, True, chart_rect, [(sample.time, sample.user + sample.sys + sample.io) for sample in cpu_stats], proc_tree) # render CPU load - draw_chart(ctx, CPU_COLOR, True, chart_rect, [(sample.time, sample.user + sample.sys) for sample in cpu_stats], proc_tree) + #draw_chart(ctx, CPU_COLOR, True, chart_rect, [(sample.time, sample.user + sample.sys) for sample in cpu_stats], proc_tree) - curr_y = curr_y + 30 + bar_h + #curr_y = curr_y + 30 + bar_h # render second chart - draw_legend_line(ctx, "Disk throughput", DISK_TPUT_COLOR, off_x, curr_y+20, leg_s) - draw_legend_box(ctx, "Disk utilization", IO_COLOR, off_x + 120, curr_y+20, leg_s) + #draw_legend_line(ctx, "Disk throughput", DISK_TPUT_COLOR, off_x, curr_y+20, leg_s) + #draw_legend_box(ctx, "Disk utilization", IO_COLOR, off_x + 120, curr_y+20, leg_s) # render I/O utilization - chart_rect = (off_x, curr_y+30, w, bar_h) - draw_box_ticks(ctx, chart_rect, sec_w) - draw_chart(ctx, IO_COLOR, True, chart_rect, [(sample.time, sample.util) for sample in disk_stats], proc_tree) + #chart_rect = (off_x, curr_y+30, w, bar_h) + #draw_box_ticks(ctx, chart_rect, sec_w) + #draw_chart(ctx, IO_COLOR, True, chart_rect, [(sample.time, sample.util) for sample in disk_stats], proc_tree) # render disk throughput - max_sample = max(disk_stats, key=lambda s: s.tput) - draw_chart(ctx, DISK_TPUT_COLOR, False, chart_rect, [(sample.time, sample.tput) for sample in disk_stats], proc_tree) + #max_sample = max(disk_stats, key=lambda s: s.tput) + #draw_chart(ctx, DISK_TPUT_COLOR, False, chart_rect, [(sample.time, sample.tput) for sample in disk_stats], proc_tree) - pos_x = off_x + ((max_sample.time - proc_tree.start_time) * w / proc_tree.duration) + #pos_x = off_x + ((max_sample.time - proc_tree.start_time) * w / proc_tree.duration) + pos_x = off_x shift_x, shift_y = -20, 20 if (pos_x < off_x + 245): shift_x, shift_y = 5, 40 - label = "%dMB/s" % round((max_sample.tput) / 1024.0) - draw_text(ctx, label, DISK_TPUT_COLOR, pos_x + shift_x, curr_y + shift_y) + #label = "%dMB/s" % round((max_sample.tput) / 1024.0) + #draw_text(ctx, label, DISK_TPUT_COLOR, pos_x + shift_x, curr_y + shift_y) - # draw process boxes - draw_process_bar_chart(ctx, proc_tree, curr_y + bar_h, w, h) - ctx.set_font_size(SIG_FONT_SIZE) - draw_text(ctx, SIGNATURE, SIG_COLOR, off_x + 5, h - off_y - 5) + chart_rect = [off_x, curr_y+60, w, h - 2 * off_y - (curr_y+60) + proc_h] + + draw_legend_box(ctx, "Configure", TASK_COLOR_CONFIGURE, off_x , curr_y + 45, leg_s) + draw_legend_box(ctx, "Compile", TASK_COLOR_COMPILE, off_x+120, curr_y + 45, leg_s) + draw_legend_box(ctx, "Install", TASK_COLOR_INSTALL, off_x+240, curr_y + 45, leg_s) + draw_legend_box(ctx, "Package", TASK_COLOR_PACKAGE, off_x+360, curr_y + 45, leg_s) + draw_legend_box(ctx, "Populate Sysroot", TASK_COLOR_SYSROOT, off_x+480, curr_y + 45, leg_s) -def draw_process_bar_chart(ctx, proc_tree, curr_y, w, h): - draw_legend_box(ctx, "Running (%cpu)", PROC_COLOR_R, off_x , curr_y + 45, leg_s) - draw_legend_box(ctx, "Unint.sleep (I/O)", PROC_COLOR_D, off_x+120, curr_y + 45, leg_s) - draw_legend_box(ctx, "Sleeping", PROC_COLOR_S, off_x+240, curr_y + 45, leg_s) - draw_legend_box(ctx, "Zombie", PROC_COLOR_Z, off_x+360, curr_y + 45, leg_s) - - chart_rect = [off_x, curr_y+60, w, h - 2 * off_y - (curr_y+60) + proc_h] ctx.set_font_size(PROC_TEXT_FONT_SIZE) draw_box_ticks(ctx, chart_rect, sec_w) draw_5sec_labels(ctx, chart_rect, sec_w) - + y = curr_y+60 + + offset = min(res.start.keys()) + for s in sorted(res.start.keys()): + task = res.start[s].split(":")[1] + #print res.start[s] + #print res.processes[res.start[s]][1] + #print s + x = (s - offset) * sec_w + w = ((res.processes[res.start[s]][1] - s) * sec_w) + + #print "proc at %s %s %s %s" % (x, y, w, proc_h) + col = None + if task == "do_compile": + col = TASK_COLOR_COMPILE + elif task == "do_configure": + col = TASK_COLOR_CONFIGURE + elif task == "do_install": + col = TASK_COLOR_INSTALL + elif task == "do_package": + col = TASK_COLOR_PACKAGE + elif task == "do_populate_sysroot": + col = TASK_COLOR_SYSROOT + + draw_rect(ctx, PROC_BORDER_COLOR, (x, y, w, proc_h)) + if col: + draw_fill_rect(ctx, col, (x, y, w, proc_h)) + + draw_label_in_box(ctx, PROC_TEXT_COLOR, res.start[s], x, y + proc_h - 4, w, proc_h) + y = y + proc_h + + # draw process boxes + #draw_process_bar_chart(ctx, proc_tree, curr_y + bar_h, w, h) + + ctx.set_font_size(SIG_FONT_SIZE) + draw_text(ctx, SIGNATURE, SIG_COLOR, off_x + 5, h - off_y - 5) + +def draw_process_bar_chart(ctx, proc_tree, curr_y, w, h): + for root in proc_tree.process_tree: draw_processes_recursively(ctx, root, proc_tree, y, proc_h, chart_rect) y = y + proc_h * proc_tree.num_nodes([root])
batch.py +2 -2
@@ -34,11 +34,11 @@ return 10 make_surface, write_surface = handlers[format] - w,h = draw.extents(*res) + w,h = draw.extents(res) w = max(w, draw.MIN_IMG_W) surface = make_surface(w,h) ctx = cairo.Context(surface) - draw.render(ctx, options, *res) + draw.render(ctx, options, res) write_surface(surface) writer.status("bootchart written to '%s'" % filename)
parsing.py +26 -28
78 lines not on this page · raw message
main.py +2 -0
@@ -102,6 +102,8 @@ res = parsing.parse(writer, args, options.prune) + + #print str(res) if options.interactive or options.output == None: gui.show(res, options)
This is a chart of pseudo-native as of master before the last few bitbake commits:
http://tim.rpsys.net/bootchart-pseudo1.png
and after the recent bitbake commits, this is the result:
http://tim.rpsys.net/bootchart-pseudo1.png
Whats interesting in the gradient of the rapid fire tasks at the start of the graph. There is now much less "free" space between the tasks and task execution and throughput was therefore improved and the gradient much sharper.
Looking at the graphs, the new task execution order is suboptimal though, mainly as I changed something there that before realising there was another problem that was making this hurt more than it should.
I think reverting:
http://git.pokylinux.org/cgit.cgi/poky/commit/?id=974ea1a190167dcfd831ba1fc5f733e0dc9a6fda
should help parallelism a little further.
I'm running a full build to see what this means in read world numbers. It certainly executed all the non-critical path tasks faster with these changes. I pulled the graph of the first 1000 tasks and they executed in a faster time period, its difficult to say what the effect on the overall build will be at this point though.
Interesting what some numbers can do to point out problems though.
Just for the record, I used a dummy recipe of empty tasks to test how fast the execution code was combined with bitbake profiling and some suitable tweaks to the looping delays to make it clear when bitbake was sleeping and shouldn't be. These changes dropped the test time from 9 seconds to 5 (where the timing includes bitbake startup and cache loading overheads).
Cheers,
Richard
Quoted 31 lines · Richard Purdie
We have just merged Beth's initial buildstats logging work. I was sitting wondering how to actually evaluate the numbers as I wanted to know "where are we spending the time?".
It occurred to me that I wanted a graph very similar to that generated by bootchart. I looked around and found pyboootchartgui and then hacked it around a bit and coerced it to start producing charts like:
http://tim.rpsys.net/bootchart.png
which is the initial "pseudo-native" part of the build. This was simple enough to test with.
I then tried graphing a poky-image-sato. To get a graph I could actually read, I stripped out any task taking less than 8 seconds and scaled the x axis from 25 units per second to one unit per second. The result was:
http://tim.rpsys.net/bootchart2.png
(warning this is a 2.7MB png)I also added in a little bit of colour coding for the second chart. Interestingly it looks like there is more yellow than green meaning configure is a bigger drain on the build time not that its unexpected :/.
I quite enjoyed playing with this and on a serious note, the gradient of the task graph makes me a little suspicious of whether the overhead of launching tasks in bitbake itself is having some effect on build time. Certainly on the first graph there are some interesting latencies showing up.
Quoted 5 lines
Anyhow, I think this is the first time bitbake's task execution has been visualised and there are some interesting things we can learn from it. I'm hoping this is a start of a much more detailed understanding of the build process with respect to performance.
Per Richard's request, I ran this on a large Westmere machine with 40 physical cores (10 per socket).
BB_NUMBER_THREADS = "40"PARALLEL_MAKE = "-j 40"
MACHINE=qemux86
Here's the poky-image-sato "buildchart" (1.6M):
http://dvhart.com/darren/yocto/buildchart-sato.png
For reference, total build time:
real 68m57.268s user 424m23.566s sys 211m30.272s
The higher core count results in a steeper gradient and makes the serialized build sections a bit more prominent. The two obvious serialization points being "eglibc" and the "openssl,ncurses,gettext,readline,sqlite,augeas,python,boost,perl" sequence.
--
Darren
Quoted 10 lines
Cheers,
Richard
_______________________________________________
poky mailing list
[email protected]
https://lists.yoctoproject.org/listinfo/poky
Signature · 3 lines
Darren Hart Intel Open Source Technology Center Yocto Project - Linux Kernel
Quoted 55 lines · Darren Hart
Quoted 31 lines · Richard Purdie
We have just merged Beth's initial buildstats logging work. I was sitting wondering how to actually evaluate the numbers as I wanted to know "where are we spending the time?".
It occurred to me that I wanted a graph very similar to that generated by bootchart. I looked around and found pyboootchartgui and then hacked it around a bit and coerced it to start producing charts like:
http://tim.rpsys.net/bootchart.png
which is the initial "pseudo-native" part of the build. This was simple enough to test with.
I then tried graphing a poky-image-sato. To get a graph I could actually read, I stripped out any task taking less than 8 seconds and scaled the x axis from 25 units per second to one unit per second. The result was:
http://tim.rpsys.net/bootchart2.png
(warning this is a 2.7MB png)I also added in a little bit of colour coding for the second chart. Interestingly it looks like there is more yellow than green meaning configure is a bigger drain on the build time not that its unexpected :/.
I quite enjoyed playing with this and on a serious note, the gradient of the task graph makes me a little suspicious of whether the overhead of launching tasks in bitbake itself is having some effect on build time. Certainly on the first graph there are some interesting latencies showing up.
Quoted 5 lines
Anyhow, I think this is the first time bitbake's task execution has been visualised and there are some interesting things we can learn from it. I'm hoping this is a start of a much more detailed understanding of the build process with respect to performance.
Per Richard's request, I ran this on a large Westmere machine with 40 physical cores (10 per socket).
BB_NUMBER_THREADS= "40"PARALLEL_MAKE= "-j 40"
MACHINE=qemux86Here's the poky-image-sato "buildchart" (1.6M):
http://dvhart.com/darren/yocto/buildchart-sato.png
For reference, total build time:
real 68m57.268s
user 424m23.566s
sys 211m30.272s
Quoted 5 lines
The higher core count results in a steeper gradient and makes the serialized build sections a bit more prominent. The two obvious serialization points being "eglibc" and the "openssl,ncurses,gettext,readline,sqlite,augeas,python,boost,perl" sequence.
Note: I am not the only one running jobs on the 40-way, but the CPU utilization has been under 30% during my periodic spot checks.
With poky/master 27894c58b74fb01530aa7144b8e8c7a2298187b4 with perl-native forced to use PARALLEL_MAKE="" we get:
http://dvhart.com/darren/yocto/buildchart-pre-perl.png
real 58m28.575s user 421m47.013s sys 193m12.428s
The second pinch point (gettext,python,etc.) seems to be less pronounced, and we're 10 minutes faster (but that shouldn't be taken too literally with other jobs running on this system).
With contrib/nitin/misc 42a0ae1220f9add4078f207086ed7f9c859de0af which includes a fix for perl-native PARALLEL_MAKE:
http://dvhart.com/darren/yocto/buildchart-post-perl.png
real 57m42.992s user 423m5.026s sys 197m50.407s
Visual inspections suggest the second pinch point is further reduced here.
--
Darren
Quoted 16 lines
--
DarrenQuoted 10 lines
Cheers,
Richard
_______________________________________________
poky mailing list
[email protected]
https://lists.yoctoproject.org/listinfo/poky
Signature · 3 lines
Darren Hart Intel Open Source Technology Center Yocto Project - Linux Kernel