Skip to contentSkip to conversation

perf report: show sort_order in title

PATCH In review [email protected] 2 participants 6 messages Mar 13, 2017 – Mar 15, 2017 perf events

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 perf report: show sort_order in title

From: Changbin Du <[email protected]>

In the report, how does the data sort is a important info for analyser. Moreover, perf sometimes insert or append sort fields automatically. Thus user may confuse how it sorts w/o reading perf internal. So here print the order info to repor title.

---
 tools/perf/ui/browsers/hists.c | 22 ++++++++++++++++++++++
 1 file changed, 22 insertions(+)
tools/perf/ui/browsers/hists.c +22 -0
index fc4fb66..6c6b615 100644
@@ -2194,6 +2194,25 @@ static inline bool is_report_browser(void *timer)
 	return timer == NULL;
 }
 
+static int get_sort_fields_str(struct hists *hists, char *buf, size_t size)
+{
+	struct perf_hpp_fmt *fmt;
+	bool first = true;
+	int ret = 0;
+
+	hists__for_each_sort_list(hists, fmt) {
+		if (first) {
+			first = false;
+			ret += scnprintf(buf + ret, size - ret, "%s", fmt->name);
+		} else {
+			ret += scnprintf(buf + ret, size - ret, ",%s", fmt->name);
+		}
+		if (size - ret <= 0)
+			break;
+	}
+	return ret;
+}
+
 static int perf_evsel_browser_title(struct hist_browser *browser,
 				char *bf, size_t size)
 {
@@ -2274,6 +2293,9 @@ static int perf_evsel_browser_title(struct hist_browser *browser,
 			printed += scnprintf(bf + printed, size - printed, " [z]");
 	}
 
+	get_sort_fields_str(hists, buf, sizeof(buf));
+	printed += scnprintf(bf + printed, size - printed, ", Sort by: %s", buf);
+
 	return printed;
 }
 
Signature · 1 line
2.7.4
#1 lore raw

Em Mon, Mar 13, 2017 at 04:36:35PM +0800, [email protected] escreveu:

Quoted 17 lines

From: Changbin Du <[email protected]>

In the report, how does the data sort is a important info for analyser. Moreover, perf sometimes insert or append sort fields automatically. Thus user may confuse how it sorts w/o reading perf internal. So here print the order info to repor title.

---
 tools/perf/ui/browsers/hists.c | 22 ++++++++++++++++++++++
 1 file changed, 22 insertions(+)
tools/perf/ui/browsers/hists.c +0 -0
index fc4fb66..6c6b615 100644

tools/perf/ui/browsers/hists.c static inline bool is_report_browser(void *timer) · line 2197

Quoted 4 lines
tools/perf/ui/browsers/hists.c +1 -0
@@ -2194,6 +2194,25 @@ static inline bool is_report_browser(void *timer)
 	return timer == NULL;
 }
 
+static int get_sort_fields_str(struct hists *hists, char *buf, size_t size)

I'll just rename this to use the tools/perf/ style for such functions, making it:

static int hists__scnprintf_sort_fields(hists, buf, size)

Quoted 33 lines
tools/perf/ui/browsers/hists.c +21 -0
+{
+	struct perf_hpp_fmt *fmt;
+	bool first = true;
+	int ret = 0;
+
+	hists__for_each_sort_list(hists, fmt) {
+		if (first) {
+			first = false;
+			ret += scnprintf(buf + ret, size - ret, "%s", fmt->name);
+		} else {
+			ret += scnprintf(buf + ret, size - ret, ",%s", fmt->name);
+		}
+		if (size - ret <= 0)
+			break;
+	}
+	return ret;
+}
+
 static int perf_evsel_browser_title(struct hist_browser *browser,
 				char *bf, size_t size)
 {
@@ -2274,6 +2293,9 @@ static int perf_evsel_browser_title(struct hist_browser *browser,
 			printed += scnprintf(bf + printed, size - printed, " [z]");
 	}
 
+	get_sort_fields_str(hists, buf, sizeof(buf));
+	printed += scnprintf(bf + printed, size - printed, ", Sort by: %s", buf);
+
 	return printed;
 }
 
Signature · 1 line
2.7.4
#2 lore raw

Em Mon, Mar 13, 2017 at 11:47:14AM -0300, Arnaldo Carvalho de Melo escreveu:

Quoted 21 lines

Em Mon, Mar 13, 2017 at 04:36:35PM +0800, [email protected] escreveu:

Quoted 20 lines

From: Changbin Du <[email protected]>

In the report, how does the data sort is a important info for analyser. Moreover, perf sometimes insert or append sort fields automatically. Thus user may confuse how it sorts w/o reading perf internal. So here print the order info to repor title.

---
 tools/perf/ui/browsers/hists.c | 22 ++++++++++++++++++++++
 1 file changed, 22 insertions(+)
tools/perf/ui/browsers/hists.c +0 -0
index fc4fb66..6c6b615 100644
@@ -2194,6 +2194,25 @@ static inline bool is_report_browser(void *timer)
 	return timer == NULL;
 }
 
Quoted 6 lines

tools/perf/ui/browsers/hists.c static inline bool is_report_browser(void *timer) · line 2197

Quoted 1 line
tools/perf/ui/browsers/hists.c +1 -0
+static int get_sort_fields_str(struct hists *hists, char *buf, size_t size)

I'll just rename this to use the tools/perf/ style for such functions, making it:

static int hists__scnprintf_sort_fields(hists, buf, size)

But then, while testing,

Before:

  $ perf report
  Samples: 405  of event 'cycles', Event count (approx.): 101733003
  Overhead  Command          Shared Object                        Symbol
    11.15%  swapper          [kernel.vmlinux]                     [k] intel_idle
     3.00%  firefox          libxul.so                            [.] 0x0000000001298b8d
     1.74%  swapper          [kernel.vmlinux]                     [k] update_blocked_averages
     1.69%  qemu-system-x86  [kernel.vmlinux]                     [k] __fget
     1.18%  swapper          [kernel.vmlinux]                     [k] update_wall_time

Tip: Save output of perf stat using: perf stat record <target workload>

After:

  $ perf report
  Samples: 405  of event 'cycles', Event count (approx.): 101733003, Sort by: Children,Overhead,Command,Shared Object,Symbol
  Overhead  Command          Shared Object                        Symbol
    11.15%  swapper          [kernel.vmlinux]                     [k] intel_idle
     3.00%  firefox          libxul.so                            [.] 0x0000000001298b8d
     1.74%  swapper          [kernel.vmlinux]                     [k] update_blocked_averages
     1.69%  qemu-system-x86  [kernel.vmlinux]                     [k] __fget
     1.18%  swapper          [kernel.vmlinux]                     [k] update_wall_time

I see now duplication of info, where is the value? Can you show the usecase in a compelling way?

- Arnaldo

#3 lore raw
Quoted 32 lines · Arnaldo Carvalho de Melo
Quoted 4 lines

I'll just rename this to use the tools/perf/ style for such functions, making it:

static int hists__scnprintf_sort_fields(hists, buf, size)

But then, while testing,

Before:

  $ perf report
  Samples: 405  of event 'cycles', Event count (approx.): 101733003
  Overhead  Command          Shared Object                        Symbol
    11.15%  swapper          [kernel.vmlinux]                     [k] intel_idle
     3.00%  firefox          libxul.so                            [.] 0x0000000001298b8d
     1.74%  swapper          [kernel.vmlinux]                     [k] update_blocked_averages
     1.69%  qemu-system-x86  [kernel.vmlinux]                     [k] __fget
     1.18%  swapper          [kernel.vmlinux]                     [k] update_wall_time

Tip: Save output of perf stat using: perf stat record <target workload>

After:

  $ perf report
  Samples: 405  of event 'cycles', Event count (approx.): 101733003, Sort by: Children,Overhead,Command,Shared Object,Symbol
  Overhead  Command          Shared Object                        Symbol
    11.15%  swapper          [kernel.vmlinux]                     [k] intel_idle
     3.00%  firefox          libxul.so                            [.] 0x0000000001298b8d
     1.74%  swapper          [kernel.vmlinux]                     [k] update_blocked_averages
     1.69%  qemu-system-x86  [kernel.vmlinux]                     [k] __fget
     1.18%  swapper          [kernel.vmlinux]                     [k] update_wall_time
Quoted 4 lines

I see now duplication of info, where is the value? Can you show the usecase in a compelling way?

- Arnaldo

Thanks for trying. The key idea is to show how does the data sort, especially the first sort key. When I use some GUI based perf tool, I can see how my data is sorted by checking the report header status. I think this is a good for browser.

You are right, the info is duplicated. I got another idea that we show a '↓' at the header string and only for the first sort key. What do you think?

$ perf report
   Samples: 405  of event 'cycles', Event count (approx.): 101733003
   ↓Overhead  Command          Shared Object                        Symbol
     11.15%  swapper          [kernel.vmlinux]                     [k] intel_idle
      3.00%  firefox          libxul.so                            [.] 0x0000000001298b8d
      1.74%  swapper          [kernel.vmlinux]                     [k] update_blocked_averages

Another idea I want to add is to support dynamic sorting. For me, I use perf to analysing entire system performance, and the data is very large. Then sometimes it take as long as ~10 minitues to read perf data. So I think if we can change sort w/o reload data will be good.

Signature · 2 lines
Thanks,
Changbin Du
#4 lore raw

Em Tue, Mar 14, 2017 at 10:16:44AM +0800, Du, Changbin escreveu:

Quoted 27 lines · Arnaldo Carvalho de Melo
Quoted 13 lines

But then, while testing,

Before:

  $ perf report
  Samples: 405  of event 'cycles', Event count (approx.): 101733003
  Overhead  Command          Shared Object                        Symbol
    11.15%  swapper          [kernel.vmlinux]                     [k] intel_idle

Tip: Save output of perf stat using: perf stat record <target workload>

After:

Quoted 7 lines
  $ perf report
  Samples: 405  of event 'cycles', Event count (approx.): 101733003, Sort by: Children,Overhead,Command,Shared Object,Symbol
  Overhead  Command          Shared Object                        Symbol
    11.15%  swapper          [kernel.vmlinux]                     [k] intel_idle

I see now duplication of info, where is the value? Can you show the usecase in a compelling way?

Thanks for trying. The key idea is to show how does the data sort, especially the first sort key. When I use some GUI based perf tool, I can see how my data is sorted by checking the report header status. I think this is a good for browser.

Quoted 6 lines

You are right, the info is duplicated. I got another idea that we show a '↓' at the header string and only for the first sort key. What do you think?

$ perf report
   Samples: 405  of event 'cycles', Event count (approx.): 101733003
   ↓Overhead  Command          Shared Object                        Symbol

this is much more compact, but you need to make it abundantly clear what you are trying to achieve by showind counter examples were what we get on that line starting with your suggested marker isn't the sort order. Otherwise even a character is one too much :-)

Quoted 8 lines
     11.15%  swapper          [kernel.vmlinux]                     [k] intel_idle
      3.00%  firefox          libxul.so                            [.] 0x0000000001298b8d
      1.74%  swapper          [kernel.vmlinux]                     [k] update_blocked_averages

Another idea I want to add is to support dynamic sorting. For me, I use perf to analysing entire system performance, and the data is very large. Then sometimes it take as long as ~10 minitues to read perf data. So I think if we can change sort w/o reload data will be good.

And in some cases it is even possible! I.e. if you haven't collapsed too much, you will not have to reprocess the file to get to the new order.

BTW, have you played with:

  perf top --hierarchy

Try it with -g and --call-graph dwarf

Also try:

  perf report --hierarchy

- Arnaldo

#5 lore raw
Quoted 31 lines · Arnaldo Carvalho de Melo

Em Tue, Mar 14, 2017 at 10:16:44AM +0800, Du, Changbin escreveu:

Quoted 29 lines · Arnaldo Carvalho de Melo
Quoted 13 lines

But then, while testing,

Before:

  $ perf report
  Samples: 405  of event 'cycles', Event count (approx.): 101733003
  Overhead  Command          Shared Object                        Symbol
    11.15%  swapper          [kernel.vmlinux]                     [k] intel_idle

Tip: Save output of perf stat using: perf stat record <target workload>

After:

Quoted 7 lines
  $ perf report
  Samples: 405  of event 'cycles', Event count (approx.): 101733003, Sort by: Children,Overhead,Command,Shared Object,Symbol
  Overhead  Command          Shared Object                        Symbol
    11.15%  swapper          [kernel.vmlinux]                     [k] intel_idle

I see now duplication of info, where is the value? Can you show the usecase in a compelling way?

Thanks for trying. The key idea is to show how does the data sort, especially the first sort key. When I use some GUI based perf tool, I can see how my data is sorted by checking the report header status. I think this is a good for browser.

You are right, the info is duplicated. I got another idea that we show a '↓' at the header string and only for the first sort key. What do you think?

Quoted 9 lines
Quoted 3 lines
$ perf report
   Samples: 405  of event 'cycles', Event count (approx.): 101733003
   ↓Overhead  Command          Shared Object                        Symbol

this is much more compact, but you need to make it abundantly clear what you are trying to achieve by showind counter examples were what we get on that line starting with your suggested marker isn't the sort order. Otherwise even a character is one too much :-)

Yes, I just want get know how does perf data sort. Because sometimes the real sort order doesn't match the '-s' option I given. In this case, I was confused about the sorting before reading into the code.

Quoted 16 lines
Quoted 8 lines
     11.15%  swapper          [kernel.vmlinux]                     [k] intel_idle
      3.00%  firefox          libxul.so                            [.] 0x0000000001298b8d
      1.74%  swapper          [kernel.vmlinux]                     [k] update_blocked_averages

Another idea I want to add is to support dynamic sorting. For me, I use perf to analysing entire system performance, and the data is very large. Then sometimes it take as long as ~10 minitues to read perf data. So I think if we can change sort w/o reload data will be good.

And in some cases it is even possible! I.e. if you haven't collapsed too much, you will not have to reprocess the file to get to the new order.

BTW, have you played with:

  perf top --hierarchy
Quoted 7 lines

Try it with -g and --call-graph dwarf

Also try:

  perf report --hierarchy

- Arnaldo

Sounds great! I tried '--hierarchy' option, but still don't know how to resort the report. Could you give a hint?

Signature · 2 lines
Thanks,
Changbin Du