Arm64 Neoverse CPUs supports data source in Arm SPE trace, this allows
us to detect cache line contention and transfers.
Unlike x86 architecture, Arm SPE trace data cannot provide 'HITM'
snooping flag, Ali Said has a patch set "perf: arm-spe: Decode SPE
source and use for perf c2c" [1] which introduces 'peer' flag and
synthesizes memory samples with this flag.
Based on patch set [1], this patch set is to finish the second half work
to consume the 'peer' flag in perf c2c tool, it adds an extra display
'peer' mode.
Patches 01, 02 and 03 are to support 'N/A' metrics for store operations.
Patches 04 and 05 adds statistics and dimensions for memory samples with
peer flag.
Patches 06, 07, 08 are for refactoring, it refines the code with more
general naming so this can allow us to easier to extend display modes
but not strictly bound to HITM tags.
Patches 09, 10 and 11 are to extend display 'peer' mode, it also updates
the document and also changes to use 'peer' mode as default mode on
Arm64 arches.
This patch set has been verified for both x86 and Arm64 memory samples.
The display result with x86 memory samples:
=================================================
Shared Data Cache Line Table
=================================================
#
# ----------- Cacheline ---------- Tot ------- Load Hitm ------- Snoop Total Total Total --------- Stores -------- ----- Core Load Hit ----- - LLC Load Hit -- - RMT Load Hit -- --- Load Dram ----
# Index Address Node PA cnt Hitm Total LclHitm RmtHitm Peer records Loads Stores L1Hit L1Miss N/A FB L1 L2 LclHit LclHitm RmtHit RmtHitm Lcl Rmt
# ..... .................. .... ...... ....... ....... ....... ....... ....... ....... ....... ....... ....... ....... ....... ....... ....... ....... ........ ....... ........ ....... ........ ........
#
0 0x55c8971f0080 0 1967 66.14% 252 252 0 0 6044 3550 2494 2024 470 0 528 2672 78 20 252 0 0 0 0
1 0x55c8971f00c0 0 1 33.86% 129 129 0 0 914 914 0 0 0 0 272 374 52 87 129 0 0 0 0
=================================================
Shared Cache Line Distribution Pareto
=================================================
#
# ----- HITM ----- Snoop ------- Store Refs ------ --------- Data address --------- --------------- cycles --------------- Total cpu Shared
# Num RmtHitm LclHitm Peer L1 Hit L1 Miss N/A Offset Node PA cnt Code address rmt hitm lcl hitm load peer records cnt Symbol Object Source:Line Node
# ..... ....... ....... ....... ....... ....... ....... .................. .... ...... .................. ........ ........ ........ ........ ....... ........ ...................... ................. ....................... ....
#
-------------------------------------------------------------------------------
0 0 252 0 2024 470 0 0x55c8971f0080
-------------------------------------------------------------------------------
0.00% 12.30% 0.00% 0.00% 0.00% 0.00% 0x0 0 1 0x55c8971ed3e9 0 1313 863 0 1222 3 [.] 0x00000000000013e9 false_sharing.exe false_sharing.exe[13e9] 0
0.00% 0.79% 0.00% 90.51% 0.00% 0.00% 0x0 0 1 0x55c8971ed3e2 0 1800 878 0 3029 3 [.] 0x00000000000013e2 false_sharing.exe false_sharing.exe[13e2] 0
0.00% 0.00% 0.00% 9.49% 100.00% 0.00% 0x0 0 1 0x55c8971ed3f4 0 0 0 0 662 3 [.] 0x00000000000013f4 false_sharing.exe false_sharing.exe[13f4] 0
0.00% 86.90% 0.00% 0.00% 0.00% 0.00% 0x20 0 1 0x55c8971ed447 0 141 103 0 1131 2 [.] 0x0000000000001447 false_sharing.exe false_sharing.exe[1447] 0
-------------------------------------------------------------------------------
1 0 129 0 0 0 0 0x55c8971f00c0
-------------------------------------------------------------------------------
0.00% 100.00% 0.00% 0.00% 0.00% 0.00% 0x20 0 1 0x55c8971ed455 0 88 94 0 914 2 [.] 0x0000000000001455 false_sharing.exe false_sharing.exe[1455] 0
The display result with Arm SPE memory samples:
=================================================
Shared Data Cache Line Table
=================================================
#
# ----------- Cacheline ---------- Snoop ------- Load Hitm ------- Snoop Total Total Total --------- Stores -------- ----- Core Load Hit ----- - LLC Load Hit -- - RMT Load Hit -- --- Load Dram ----
# Index Address Node PA cnt Peer Total LclHitm RmtHitm Peer records Loads Stores L1Hit L1Miss N/A FB L1 L2 LclHit LclHitm RmtHit RmtHitm Lcl Rmt
# ..... .................. .... ...... ....... ....... ....... ....... ....... ....... ....... ....... ....... ....... ....... ....... ....... ....... ........ ....... ........ ....... ........ ........
#
0 0xaaaac17d6000 N/A 0 100.00% 0 0 0 99 18851 18851 0 0 0 0 0 18752 0 0 0 0 0 0 0
=================================================
Shared Cache Line Distribution Pareto
=================================================
#
# ----- HITM ----- Snoop ------- Store Refs ------ --------- Data address --------- --------------- cycles --------------- Total cpu Shared
# Num RmtHitm LclHitm Peer L1 Hit L1 Miss N/A Offset Node PA cnt Code address rmt hitm lcl hitm load peer records cnt Symbol Object Source:Line Node
# ..... ....... ....... ....... ....... ....... ....... .................. .... ...... .................. ........ ........ ........ ........ ....... ........ ...................... ................ ............... ....
#
-------------------------------------------------------------------------------
0 0 0 99 0 0 0 0xaaaac17d6000
-------------------------------------------------------------------------------
0.00% 0.00% 6.06% 0.00% 0.00% 0.00% 0x20 N/A 0 0xaaaac17c25ac 0 0 43 375 18469 2 [.] 0x00000000000025ac memstress memstress[25ac] 0
0.00% 0.00% 93.94% 0.00% 0.00% 0.00% 0x29 N/A 0 0xaaaac17c3e88 0 0 173 180 135 2 [.] 0x0000000000003e88 memstress memstress[3e88] 0
[1] https://lore.kernel.org/lkml/[email protected]/
Changes from v1:
* Update patches 01, 02 and 03 to support 'N/A' metrics for store
operations, so can align with the patch set [1] for store samples.
Leo Yan (11):
perf mem: Add stats for store operation with no available memory level
perf c2c: Add dimensions for 'N/A' metrics of store operation
perf c2c: Update documentation for store metric 'N/A'
perf mem: Add statistics for peer snooping
perf c2c: Add dimensions for peer load operations
perf c2c: Use explicit names for display macros
perf c2c: Rename dimension from 'percent_hitm' to
'percent_costly_snoop'
perf c2c: Refactor node header
perf c2c: Sort on peer snooping for load operations
perf c2c: Update documentation for new display option 'peer'
perf c2c: Use 'peer' as default display for Arm64
tools/perf/Documentation/perf-c2c.txt | 30 ++-
tools/perf/builtin-c2c.c | 363 ++++++++++++++++++++------
tools/perf/util/mem-events.c | 14 +-
tools/perf/util/mem-events.h | 2 +
4 files changed, 323 insertions(+), 86 deletions(-)
--
2.25.1
Use more general naming for the main sort dimension, this can allow us
not to sort only on HITM snoop type, so it can be extended to support
other costly snooping operations. So rename the dimension to the prefix
'percent_costly_".
Signed-off-by: Leo Yan <[email protected]>
Tested-by: Ali Saidi <[email protected]>
---
tools/perf/builtin-c2c.c | 40 ++++++++++++++++++++--------------------
1 file changed, 20 insertions(+), 20 deletions(-)
diff --git a/tools/perf/builtin-c2c.c b/tools/perf/builtin-c2c.c
index 180208a4085f..1e1565e5d396 100644
--- a/tools/perf/builtin-c2c.c
+++ b/tools/perf/builtin-c2c.c
@@ -794,7 +794,7 @@ percent_color(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
return hpp_color_scnprintf(hpp, "%*.2f%%", width - 1, per);
}
-static double percent_hitm(struct c2c_hist_entry *c2c_he)
+static double percent_costly_snoop(struct c2c_hist_entry *c2c_he)
{
struct c2c_hists *hists;
struct c2c_stats *stats;
@@ -834,8 +834,8 @@ static double percent_hitm(struct c2c_hist_entry *c2c_he)
})
static int
-percent_hitm_entry(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
- struct hist_entry *he)
+percent_costly_snoop_entry(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
+ struct hist_entry *he)
{
struct c2c_hist_entry *c2c_he;
int width = c2c_width(fmt, hpp, he->hists);
@@ -843,20 +843,20 @@ percent_hitm_entry(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
double per;
c2c_he = container_of(he, struct c2c_hist_entry, he);
- per = percent_hitm(c2c_he);
+ per = percent_costly_snoop(c2c_he);
return scnprintf(hpp->buf, hpp->size, "%*s", width, PERC_STR(buf, per));
}
static int
-percent_hitm_color(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
- struct hist_entry *he)
+percent_costly_snoop_color(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
+ struct hist_entry *he)
{
- return percent_color(fmt, hpp, he, percent_hitm);
+ return percent_color(fmt, hpp, he, percent_costly_snoop);
}
static int64_t
-percent_hitm_cmp(struct perf_hpp_fmt *fmt __maybe_unused,
- struct hist_entry *left, struct hist_entry *right)
+percent_costly_snoop_cmp(struct perf_hpp_fmt *fmt __maybe_unused,
+ struct hist_entry *left, struct hist_entry *right)
{
struct c2c_hist_entry *c2c_left;
struct c2c_hist_entry *c2c_right;
@@ -866,8 +866,8 @@ percent_hitm_cmp(struct perf_hpp_fmt *fmt __maybe_unused,
c2c_left = container_of(left, struct c2c_hist_entry, he);
c2c_right = container_of(right, struct c2c_hist_entry, he);
- per_left = percent_hitm(c2c_left);
- per_right = percent_hitm(c2c_right);
+ per_left = percent_costly_snoop(c2c_left);
+ per_right = percent_costly_snoop(c2c_right);
return per_left - per_right;
}
@@ -1544,17 +1544,17 @@ static struct c2c_dimension dim_tot_loads = {
.width = 7,
};
-static struct c2c_header percent_hitm_header[] = {
+static struct c2c_header percent_costly_snoop_header[] = {
[DISPLAY_LCL_HITM] = HEADER_BOTH("Lcl", "Hitm"),
[DISPLAY_RMT_HITM] = HEADER_BOTH("Rmt", "Hitm"),
[DISPLAY_TOT_HITM] = HEADER_BOTH("Tot", "Hitm"),
};
-static struct c2c_dimension dim_percent_hitm = {
- .name = "percent_hitm",
- .cmp = percent_hitm_cmp,
- .entry = percent_hitm_entry,
- .color = percent_hitm_color,
+static struct c2c_dimension dim_percent_costly_snoop = {
+ .name = "percent_costly_snoop",
+ .cmp = percent_costly_snoop_cmp,
+ .entry = percent_costly_snoop_entry,
+ .color = percent_costly_snoop_color,
.width = 7,
};
@@ -1763,7 +1763,7 @@ static struct c2c_dimension *dimensions[] = {
&dim_ld_rmthit,
&dim_tot_recs,
&dim_tot_loads,
- &dim_percent_hitm,
+ &dim_percent_costly_snoop,
&dim_percent_rmt_hitm,
&dim_percent_lcl_hitm,
&dim_percent_ld_peer,
@@ -2665,7 +2665,7 @@ static int ui_quirks(void)
nodestr = "CL";
}
- dim_percent_hitm.header = percent_hitm_header[c2c.display];
+ dim_percent_costly_snoop.header = percent_costly_snoop_header[c2c.display];
/* Fix the zero line for dcacheline column. */
buf = fill_line("Cacheline", dim_dcacheline.width +
@@ -2993,7 +2993,7 @@ static int perf_c2c__report(int argc, const char **argv)
"dcacheline,"
"dcacheline_node,"
"dcacheline_count,"
- "percent_hitm,"
+ "percent_costly_snoop,"
"tot_hitm,lcl_hitm,rmt_hitm,"
"ld_peer,"
"tot_recs,"
--
2.25.1
Since Arm64 arch doesn't support HITMs flags, so use 'peer' as default
if user doesn't specify display type; for other arches, it still uses
'tot' as default display type if user doesn't specify it.
Suggested-by: Ali Saidi <[email protected]>
Signed-off-by: Leo Yan <[email protected]>
---
tools/perf/builtin-c2c.c | 21 +++++++++++++++++----
1 file changed, 17 insertions(+), 4 deletions(-)
diff --git a/tools/perf/builtin-c2c.c b/tools/perf/builtin-c2c.c
index 9b6d2a9553e1..774da4a60c08 100644
--- a/tools/perf/builtin-c2c.c
+++ b/tools/perf/builtin-c2c.c
@@ -2787,7 +2787,7 @@ static int setup_callchain(struct evlist *evlist)
static int setup_display(const char *str)
{
- const char *display = str ?: "tot";
+ const char *display = str;
if (!strcmp(display, "tot"))
c2c.display = DISPLAY_TOT_HITM;
@@ -2973,9 +2973,6 @@ static int perf_c2c__report(int argc, const char **argv)
data.path = input_name;
data.force = symbol_conf.force;
- err = setup_display(display);
- if (err)
- goto out;
err = setup_coalesce(coalesce, no_source);
if (err) {
@@ -2996,6 +2993,22 @@ static int perf_c2c__report(int argc, const char **argv)
goto out;
}
+ /*
+ * Use the 'tot' as default display type if user doesn't specify it;
+ * since Arm64 platform doesn't support HITMs flag, use 'peer' as the
+ * default display type.
+ */
+ if (!display) {
+ if (!strcmp(perf_env__arch(&session->header.env), "arm64"))
+ display = "peer";
+ else
+ display = "tot";
+ }
+
+ err = setup_display(display);
+ if (err)
+ goto out_session;
+
session->itrace_synth_opts = &itrace_synth_opts;
err = setup_nodes(session);
--
2.25.1
Sometimes we don't know memory store operations happen on exactly which
memory (or cache) level, the memory level flag is set to PERF_MEM_LVL_NA
in this case; a practical example is Arm SPE AUX trace sets this flag
for all store operations due to absent info for cache level.
This patch is to add a new item "st_na" in structure c2c_stats to add
statistics for store operations with no available cache level.
Signed-off-by: Leo Yan <[email protected]>
---
tools/perf/util/mem-events.c | 3 +++
tools/perf/util/mem-events.h | 1 +
2 files changed, 4 insertions(+)
diff --git a/tools/perf/util/mem-events.c b/tools/perf/util/mem-events.c
index db5225caaabe..5dca1882c284 100644
--- a/tools/perf/util/mem-events.c
+++ b/tools/perf/util/mem-events.c
@@ -621,6 +621,8 @@ do { \
}
if (lvl & P(LVL, MISS))
if (lvl & P(LVL, L1)) stats->st_l1miss++;
+ if (lvl & P(LVL, NA))
+ stats->st_na++;
} else {
/* unparsable data_src? */
stats->noparse++;
@@ -647,6 +649,7 @@ void c2c_add_stats(struct c2c_stats *stats, struct c2c_stats *add)
stats->st_noadrs += add->st_noadrs;
stats->st_l1hit += add->st_l1hit;
stats->st_l1miss += add->st_l1miss;
+ stats->st_na += add->st_na;
stats->load += add->load;
stats->ld_excl += add->ld_excl;
stats->ld_shared += add->ld_shared;
diff --git a/tools/perf/util/mem-events.h b/tools/perf/util/mem-events.h
index 916242f8020a..8a8b568baeee 100644
--- a/tools/perf/util/mem-events.h
+++ b/tools/perf/util/mem-events.h
@@ -63,6 +63,7 @@ struct c2c_stats {
u32 st_noadrs; /* cacheable store with no address */
u32 st_l1hit; /* count of stores that hit L1D */
u32 st_l1miss; /* count of stores that miss L1D */
+ u32 st_na; /* count of stores with memory level is not available */
u32 load; /* count of all loads in trace */
u32 ld_excl; /* exclusive loads, rmt/lcl DRAM - snp none/miss */
u32 ld_shared; /* shared loads, rmt/lcl DRAM - snp hit */
--
2.25.1
Since we have added the flag PERF_MEM_SNOOPX_PEER to support cache
snooping from peer core or cluster, this patch is to add statistics for
this new flag.
Signed-off-by: Leo Yan <[email protected]>
Tested-by: Ali Saidi <[email protected]>
---
tools/perf/util/mem-events.c | 11 ++++++++++-
tools/perf/util/mem-events.h | 1 +
2 files changed, 11 insertions(+), 1 deletion(-)
diff --git a/tools/perf/util/mem-events.c b/tools/perf/util/mem-events.c
index 5dca1882c284..6cd2b6ffb4a4 100644
--- a/tools/perf/util/mem-events.c
+++ b/tools/perf/util/mem-events.c
@@ -525,6 +525,7 @@ int c2c_decode_stats(struct c2c_stats *stats, struct mem_info *mi)
u64 op = data_src->mem_op;
u64 lvl = data_src->mem_lvl;
u64 snoop = data_src->mem_snoop;
+ u64 snoopx = data_src->mem_snoopx;
u64 lock = data_src->mem_lock;
u64 blk = data_src->mem_blk;
/*
@@ -567,10 +568,17 @@ do { \
if (lvl & P(LVL, IO)) stats->ld_io++;
if (lvl & P(LVL, LFB)) stats->ld_fbhit++;
if (lvl & P(LVL, L1 )) stats->ld_l1hit++;
- if (lvl & P(LVL, L2 )) stats->ld_l2hit++;
+ if (lvl & P(LVL, L2)) {
+ if (snoopx & P(SNOOPX, PEER))
+ stats->ld_peer++;
+ else
+ stats->ld_l2hit++;
+ }
if (lvl & P(LVL, L3 )) {
if (snoop & P(SNOOP, HITM))
HITM_INC(lcl_hitm);
+ else if (snoopx & P(SNOOPX, PEER))
+ stats->ld_peer++;
else
stats->ld_llchit++;
}
@@ -661,6 +669,7 @@ void c2c_add_stats(struct c2c_stats *stats, struct c2c_stats *add)
stats->ld_l1hit += add->ld_l1hit;
stats->ld_l2hit += add->ld_l2hit;
stats->ld_llchit += add->ld_llchit;
+ stats->ld_peer += add->ld_peer;
stats->lcl_hitm += add->lcl_hitm;
stats->rmt_hitm += add->rmt_hitm;
stats->tot_hitm += add->tot_hitm;
diff --git a/tools/perf/util/mem-events.h b/tools/perf/util/mem-events.h
index 8a8b568baeee..4879b841c841 100644
--- a/tools/perf/util/mem-events.h
+++ b/tools/perf/util/mem-events.h
@@ -75,6 +75,7 @@ struct c2c_stats {
u32 ld_l1hit; /* count of loads that hit L1D */
u32 ld_l2hit; /* count of loads that hit L2D */
u32 ld_llchit; /* count of loads that hit LLC */
+ u32 ld_peer; /* count of loads that hit peer core or cluster cache */
u32 lcl_hitm; /* count of loads with local HITM */
u32 rmt_hitm; /* count of loads with remote HITM */
u32 tot_hitm; /* count of loads with local and remote HITM */
--
2.25.1
Since now we have the statistics 'st_na' for store operations, add
dimensions for the 'N/A' (no available memory level) metrics and the
associated percentage calculation for the single cache line view.
Signed-off-by: Leo Yan <[email protected]>
---
tools/perf/builtin-c2c.c | 80 ++++++++++++++++++++++++++++++++++++----
1 file changed, 73 insertions(+), 7 deletions(-)
diff --git a/tools/perf/builtin-c2c.c b/tools/perf/builtin-c2c.c
index fbbed434014f..c8230c48125f 100644
--- a/tools/perf/builtin-c2c.c
+++ b/tools/perf/builtin-c2c.c
@@ -653,6 +653,7 @@ STAT_FN(lcl_hitm)
STAT_FN(store)
STAT_FN(st_l1hit)
STAT_FN(st_l1miss)
+STAT_FN(st_na)
STAT_FN(ld_fbhit)
STAT_FN(ld_l1hit)
STAT_FN(ld_l2hit)
@@ -677,7 +678,8 @@ static uint64_t total_records(struct c2c_stats *stats)
total = ldcnt +
stats->st_l1hit +
- stats->st_l1miss;
+ stats->st_l1miss +
+ stats->st_na;
return total;
}
@@ -899,6 +901,7 @@ PERCENT_FN(rmt_hitm)
PERCENT_FN(lcl_hitm)
PERCENT_FN(st_l1hit)
PERCENT_FN(st_l1miss)
+PERCENT_FN(st_na)
static int
percent_rmt_hitm_entry(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
@@ -1024,6 +1027,37 @@ percent_stores_l1miss_cmp(struct perf_hpp_fmt *fmt __maybe_unused,
return per_left - per_right;
}
+static int
+percent_stores_na_entry(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
+ struct hist_entry *he)
+{
+ int width = c2c_width(fmt, hpp, he->hists);
+ double per = PERCENT(he, st_na);
+ char buf[10];
+
+ return scnprintf(hpp->buf, hpp->size, "%*s", width, PERC_STR(buf, per));
+}
+
+static int
+percent_stores_na_color(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
+ struct hist_entry *he)
+{
+ return percent_color(fmt, hpp, he, percent_st_na);
+}
+
+static int64_t
+percent_stores_na_cmp(struct perf_hpp_fmt *fmt __maybe_unused,
+ struct hist_entry *left, struct hist_entry *right)
+{
+ double per_left;
+ double per_right;
+
+ per_left = PERCENT(left, st_na);
+ per_right = PERCENT(right, st_na);
+
+ return per_left - per_right;
+}
+
STAT_FN(lcl_dram)
STAT_FN(rmt_dram)
@@ -1351,7 +1385,7 @@ static struct c2c_dimension dim_tot_stores = {
};
static struct c2c_dimension dim_stores_l1hit = {
- .header = HEADER_SPAN("---- Stores ----", "L1Hit", 1),
+ .header = HEADER_SPAN("--------- Stores --------", "L1Hit", 2),
.name = "stores_l1hit",
.cmp = st_l1hit_cmp,
.entry = st_l1hit_entry,
@@ -1366,8 +1400,16 @@ static struct c2c_dimension dim_stores_l1miss = {
.width = 7,
};
+static struct c2c_dimension dim_stores_na = {
+ .header = HEADER_SPAN_LOW("N/A"),
+ .name = "stores_na",
+ .cmp = st_na_cmp,
+ .entry = st_na_entry,
+ .width = 7,
+};
+
static struct c2c_dimension dim_cl_stores_l1hit = {
- .header = HEADER_SPAN("-- Store Refs --", "L1 Hit", 1),
+ .header = HEADER_SPAN("------- Store Refs ------", "L1 Hit", 2),
.name = "cl_stores_l1hit",
.cmp = st_l1hit_cmp,
.entry = st_l1hit_entry,
@@ -1382,6 +1424,14 @@ static struct c2c_dimension dim_cl_stores_l1miss = {
.width = 7,
};
+static struct c2c_dimension dim_cl_stores_na = {
+ .header = HEADER_SPAN_LOW("N/A"),
+ .name = "cl_stores_na",
+ .cmp = st_na_cmp,
+ .entry = st_na_entry,
+ .width = 7,
+};
+
static struct c2c_dimension dim_ld_fbhit = {
.header = HEADER_SPAN("----- Core Load Hit -----", "FB", 2),
.name = "ld_fbhit",
@@ -1471,7 +1521,7 @@ static struct c2c_dimension dim_percent_lcl_hitm = {
};
static struct c2c_dimension dim_percent_stores_l1hit = {
- .header = HEADER_SPAN("-- Store Refs --", "L1 Hit", 1),
+ .header = HEADER_SPAN("------- Store Refs ------", "L1 Hit", 2),
.name = "percent_stores_l1hit",
.cmp = percent_stores_l1hit_cmp,
.entry = percent_stores_l1hit_entry,
@@ -1488,6 +1538,15 @@ static struct c2c_dimension dim_percent_stores_l1miss = {
.width = 7,
};
+static struct c2c_dimension dim_percent_stores_na = {
+ .header = HEADER_SPAN_LOW("N/A"),
+ .name = "percent_stores_na",
+ .cmp = percent_stores_na_cmp,
+ .entry = percent_stores_na_entry,
+ .color = percent_stores_na_color,
+ .width = 7,
+};
+
static struct c2c_dimension dim_dram_lcl = {
.header = HEADER_SPAN("--- Load Dram ----", "Lcl", 1),
.name = "dram_lcl",
@@ -1618,8 +1677,10 @@ static struct c2c_dimension *dimensions[] = {
&dim_tot_stores,
&dim_stores_l1hit,
&dim_stores_l1miss,
+ &dim_stores_na,
&dim_cl_stores_l1hit,
&dim_cl_stores_l1miss,
+ &dim_cl_stores_na,
&dim_ld_fbhit,
&dim_ld_l1hit,
&dim_ld_l2hit,
@@ -1632,6 +1693,7 @@ static struct c2c_dimension *dimensions[] = {
&dim_percent_lcl_hitm,
&dim_percent_stores_l1hit,
&dim_percent_stores_l1miss,
+ &dim_percent_stores_na,
&dim_dram_lcl,
&dim_dram_rmt,
&dim_pid,
@@ -2149,6 +2211,7 @@ static void print_c2c__display_stats(FILE *out)
fprintf(out, " Store - no mapping : %10d\n", stats->st_noadrs);
fprintf(out, " Store L1D Hit : %10d\n", stats->st_l1hit);
fprintf(out, " Store L1D Miss : %10d\n", stats->st_l1miss);
+ fprintf(out, " Store No available memory level : %10d\n", stats->st_na);
fprintf(out, " No Page Map Rejects : %10d\n", stats->nomap);
fprintf(out, " Unable to parse data source : %10d\n", stats->noparse);
}
@@ -2171,6 +2234,7 @@ static void print_shared_cacheline_info(FILE *out)
fprintf(out, " Blocked Access on shared lines : %10d\n", stats->blk_data + stats->blk_addr);
fprintf(out, " Store HITs on shared lines : %10d\n", stats->store);
fprintf(out, " Store L1D hits on shared lines : %10d\n", stats->st_l1hit);
+ fprintf(out, " Store No available memory level : %10d\n", stats->st_na);
fprintf(out, " Total Merged records : %10d\n", hitm_cnt + stats->store);
}
@@ -2193,10 +2257,10 @@ static void print_cacheline(struct c2c_hists *c2c_hists,
fprintf(out, "\n");
}
- fprintf(out, " -------------------------------------------------------------\n");
+ fprintf(out, " ----------------------------------------------------------------------\n");
__hist_entry__snprintf(he_cl, &hpp, hpp_list);
fprintf(out, "%s\n", bf);
- fprintf(out, " -------------------------------------------------------------\n");
+ fprintf(out, " ----------------------------------------------------------------------\n");
hists__fprintf(&c2c_hists->hists, false, 0, 0, 0, out, false);
}
@@ -2213,6 +2277,7 @@ static void print_pareto(FILE *out)
"cl_lcl_hitm,"
"cl_stores_l1hit,"
"cl_stores_l1miss,"
+ "cl_stores_na,"
"dcacheline";
perf_hpp_list__init(&hpp_list);
@@ -2664,6 +2729,7 @@ static int build_cl_output(char *cl_sort, bool no_source)
"percent_lcl_hitm,"
"percent_stores_l1hit,"
"percent_stores_l1miss,"
+ "percent_stores_na,"
"offset,offset_node,dcacheline_count,",
add_pid ? "pid," : "",
add_tid ? "tid," : "",
@@ -2850,7 +2916,7 @@ static int perf_c2c__report(int argc, const char **argv)
"tot_recs,"
"tot_loads,"
"tot_stores,"
- "stores_l1hit,stores_l1miss,"
+ "stores_l1hit,stores_l1miss,stores_na,"
"ld_fbhit,ld_l1hit,ld_l2hit,"
"ld_lclhit,lcl_hitm,"
"ld_rmthit,rmt_hitm,"
--
2.25.1
This patch is to add dimensions for peer load operations, include a
dimension for the total statistics for metric 'ld_peer', and also add
dimensions for the single cache line view.
Same as HTIM metrics, this patch also adds the dimension for mean value
for peer load operations.
Signed-off-by: Leo Yan <[email protected]>
Tested-by: Ali Saidi <[email protected]>
---
tools/perf/builtin-c2c.c | 93 +++++++++++++++++++++++++++++++++++++---
1 file changed, 88 insertions(+), 5 deletions(-)
diff --git a/tools/perf/builtin-c2c.c b/tools/perf/builtin-c2c.c
index c8230c48125f..baa003c150b1 100644
--- a/tools/perf/builtin-c2c.c
+++ b/tools/perf/builtin-c2c.c
@@ -55,6 +55,7 @@ struct c2c_hists {
struct compute_stats {
struct stats lcl_hitm;
struct stats rmt_hitm;
+ struct stats ld_peer;
struct stats load;
};
@@ -154,6 +155,7 @@ static void *c2c_he_zalloc(size_t size)
init_stats(&c2c_he->cstats.lcl_hitm);
init_stats(&c2c_he->cstats.rmt_hitm);
+ init_stats(&c2c_he->cstats.ld_peer);
init_stats(&c2c_he->cstats.load);
return &c2c_he->he;
@@ -253,6 +255,8 @@ static void compute_stats(struct c2c_hist_entry *c2c_he,
update_stats(&cstats->rmt_hitm, weight);
else if (stats->lcl_hitm)
update_stats(&cstats->lcl_hitm, weight);
+ else if (stats->ld_peer)
+ update_stats(&cstats->ld_peer, weight);
else if (stats->load)
update_stats(&cstats->load, weight);
}
@@ -658,6 +662,7 @@ STAT_FN(ld_fbhit)
STAT_FN(ld_l1hit)
STAT_FN(ld_l2hit)
STAT_FN(ld_llchit)
+STAT_FN(ld_peer)
STAT_FN(rmt_hit)
static uint64_t total_records(struct c2c_stats *stats)
@@ -674,7 +679,8 @@ static uint64_t total_records(struct c2c_stats *stats)
stats->ld_l1hit +
stats->ld_l2hit +
stats->ld_llchit +
- stats->lcl_hitm;
+ stats->lcl_hitm +
+ stats->ld_peer;
total = ldcnt +
stats->st_l1hit +
@@ -730,7 +736,8 @@ static uint64_t total_loads(struct c2c_stats *stats)
stats->ld_l1hit +
stats->ld_l2hit +
stats->ld_llchit +
- stats->lcl_hitm;
+ stats->lcl_hitm +
+ stats->ld_peer;
return ldcnt;
}
@@ -899,6 +906,7 @@ static double percent_ ## __f(struct c2c_hist_entry *c2c_he) \
PERCENT_FN(rmt_hitm)
PERCENT_FN(lcl_hitm)
+PERCENT_FN(ld_peer)
PERCENT_FN(st_l1hit)
PERCENT_FN(st_l1miss)
PERCENT_FN(st_na)
@@ -965,6 +973,37 @@ percent_lcl_hitm_cmp(struct perf_hpp_fmt *fmt __maybe_unused,
return per_left - per_right;
}
+static int
+percent_ld_peer_entry(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
+ struct hist_entry *he)
+{
+ int width = c2c_width(fmt, hpp, he->hists);
+ double per = PERCENT(he, ld_peer);
+ char buf[10];
+
+ return scnprintf(hpp->buf, hpp->size, "%*s", width, PERC_STR(buf, per));
+}
+
+static int
+percent_ld_peer_color(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
+ struct hist_entry *he)
+{
+ return percent_color(fmt, hpp, he, percent_ld_peer);
+}
+
+static int64_t
+percent_ld_peer_cmp(struct perf_hpp_fmt *fmt __maybe_unused,
+ struct hist_entry *left, struct hist_entry *right)
+{
+ double per_left;
+ double per_right;
+
+ per_left = PERCENT(left, ld_peer);
+ per_right = PERCENT(right, ld_peer);
+
+ return per_left - per_right;
+}
+
static int
percent_stores_l1hit_entry(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
struct hist_entry *he)
@@ -1213,6 +1252,7 @@ __func(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp, struct hist_entry *he) \
MEAN_ENTRY(mean_rmt_entry, rmt_hitm);
MEAN_ENTRY(mean_lcl_entry, lcl_hitm);
MEAN_ENTRY(mean_load_entry, load);
+MEAN_ENTRY(mean_peer_entry, ld_peer);
static int
cpucnt_entry(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
@@ -1360,6 +1400,14 @@ static struct c2c_dimension dim_rmt_hitm = {
.width = 7,
};
+static struct c2c_dimension dim_ld_peer = {
+ .header = HEADER_BOTH("Snoop", "Peer"),
+ .name = "ld_peer",
+ .cmp = ld_peer_cmp,
+ .entry = ld_peer_entry,
+ .width = 7,
+};
+
static struct c2c_dimension dim_cl_rmt_hitm = {
.header = HEADER_SPAN("----- HITM -----", "Rmt", 1),
.name = "cl_rmt_hitm",
@@ -1376,6 +1424,14 @@ static struct c2c_dimension dim_cl_lcl_hitm = {
.width = 7,
};
+static struct c2c_dimension dim_cl_ld_peer = {
+ .header = HEADER_BOTH("Snoop", "Peer"),
+ .name = "cl_ld_peer",
+ .cmp = ld_peer_cmp,
+ .entry = ld_peer_entry,
+ .width = 7,
+};
+
static struct c2c_dimension dim_tot_stores = {
.header = HEADER_BOTH("Total", "Stores"),
.name = "tot_stores",
@@ -1520,6 +1576,15 @@ static struct c2c_dimension dim_percent_lcl_hitm = {
.width = 7,
};
+static struct c2c_dimension dim_percent_ld_peer = {
+ .header = HEADER_BOTH("Snoop", "Peer"),
+ .name = "percent_ld_peer",
+ .cmp = percent_ld_peer_cmp,
+ .entry = percent_ld_peer_entry,
+ .color = percent_ld_peer_color,
+ .width = 7,
+};
+
static struct c2c_dimension dim_percent_stores_l1hit = {
.header = HEADER_SPAN("------- Store Refs ------", "L1 Hit", 2),
.name = "percent_stores_l1hit",
@@ -1602,7 +1667,7 @@ static struct c2c_dimension dim_node = {
};
static struct c2c_dimension dim_mean_rmt = {
- .header = HEADER_SPAN("---------- cycles ----------", "rmt hitm", 2),
+ .header = HEADER_SPAN("--------------- cycles ---------------", "rmt hitm", 3),
.name = "mean_rmt",
.cmp = empty_cmp,
.entry = mean_rmt_entry,
@@ -1625,6 +1690,14 @@ static struct c2c_dimension dim_mean_load = {
.width = 8,
};
+static struct c2c_dimension dim_mean_peer = {
+ .header = HEADER_SPAN_LOW("peer"),
+ .name = "mean_peer",
+ .cmp = empty_cmp,
+ .entry = mean_peer_entry,
+ .width = 8,
+};
+
static struct c2c_dimension dim_cpucnt = {
.header = HEADER_BOTH("cpu", "cnt"),
.name = "cpucnt",
@@ -1672,8 +1745,10 @@ static struct c2c_dimension *dimensions[] = {
&dim_tot_hitm,
&dim_lcl_hitm,
&dim_rmt_hitm,
+ &dim_ld_peer,
&dim_cl_lcl_hitm,
&dim_cl_rmt_hitm,
+ &dim_cl_ld_peer,
&dim_tot_stores,
&dim_stores_l1hit,
&dim_stores_l1miss,
@@ -1691,6 +1766,7 @@ static struct c2c_dimension *dimensions[] = {
&dim_percent_hitm,
&dim_percent_rmt_hitm,
&dim_percent_lcl_hitm,
+ &dim_percent_ld_peer,
&dim_percent_stores_l1hit,
&dim_percent_stores_l1miss,
&dim_percent_stores_na,
@@ -1704,6 +1780,7 @@ static struct c2c_dimension *dimensions[] = {
&dim_mean_rmt,
&dim_mean_lcl,
&dim_mean_load,
+ &dim_mean_peer,
&dim_cpucnt,
&dim_srcline,
&dim_dcacheline_idx,
@@ -2192,6 +2269,7 @@ static void print_c2c__display_stats(FILE *out)
fprintf(out, " Load L1D hit : %10d\n", stats->ld_l1hit);
fprintf(out, " Load L2D hit : %10d\n", stats->ld_l2hit);
fprintf(out, " Load LLC hit : %10d\n", stats->ld_llchit + stats->lcl_hitm);
+ fprintf(out, " Load HIT Peer : %10d\n", stats->ld_peer);
fprintf(out, " Load Local HITM : %10d\n", stats->lcl_hitm);
fprintf(out, " Load Remote HITM : %10d\n", stats->rmt_hitm);
fprintf(out, " Load Remote HIT : %10d\n", stats->rmt_hit);
@@ -2229,6 +2307,7 @@ static void print_shared_cacheline_info(FILE *out)
fprintf(out, " Fill Buffer Hits on shared lines : %10d\n", stats->ld_fbhit);
fprintf(out, " L1D hits on shared lines : %10d\n", stats->ld_l1hit);
fprintf(out, " L2D hits on shared lines : %10d\n", stats->ld_l2hit);
+ fprintf(out, " Load HITs on peer cache lines : %10d\n", stats->ld_peer);
fprintf(out, " LLC hits on shared lines : %10d\n", stats->ld_llchit + stats->lcl_hitm);
fprintf(out, " Locked Access on shared lines : %10d\n", stats->locks);
fprintf(out, " Blocked Access on shared lines : %10d\n", stats->blk_data + stats->blk_addr);
@@ -2257,10 +2336,10 @@ static void print_cacheline(struct c2c_hists *c2c_hists,
fprintf(out, "\n");
}
- fprintf(out, " ----------------------------------------------------------------------\n");
+ fprintf(out, " -------------------------------------------------------------------------------\n");
__hist_entry__snprintf(he_cl, &hpp, hpp_list);
fprintf(out, "%s\n", bf);
- fprintf(out, " ----------------------------------------------------------------------\n");
+ fprintf(out, " -------------------------------------------------------------------------------\n");
hists__fprintf(&c2c_hists->hists, false, 0, 0, 0, out, false);
}
@@ -2275,6 +2354,7 @@ static void print_pareto(FILE *out)
cl_output = "cl_num,"
"cl_rmt_hitm,"
"cl_lcl_hitm,"
+ "cl_ld_peer,"
"cl_stores_l1hit,"
"cl_stores_l1miss,"
"cl_stores_na,"
@@ -2727,6 +2807,7 @@ static int build_cl_output(char *cl_sort, bool no_source)
c2c.use_stdio ? "cl_num_empty," : "",
"percent_rmt_hitm,"
"percent_lcl_hitm,"
+ "percent_ld_peer,"
"percent_stores_l1hit,"
"percent_stores_l1miss,"
"percent_stores_na,"
@@ -2737,6 +2818,7 @@ static int build_cl_output(char *cl_sort, bool no_source)
"mean_rmt,"
"mean_lcl,"
"mean_load,"
+ "mean_peer,"
"tot_recs,"
"cpucnt,",
add_sym ? "symbol," : "",
@@ -2913,6 +2995,7 @@ static int perf_c2c__report(int argc, const char **argv)
"dcacheline_count,"
"percent_hitm,"
"tot_hitm,lcl_hitm,rmt_hitm,"
+ "ld_peer,"
"tot_recs,"
"tot_loads,"
"tot_stores,"
--
2.25.1
Except the existed three display options 'tot', 'rmt', 'lcl', this patch
adds a new option 'peer' so can sort on the cache hit for peer snooping.
For displaying with option 'peer', the "Shared Data Cache Line Table" and
"Shared Cache Line Distribution Pareto" both sort with the metrics
"ld_peer". As result, we can get the 'peer' display as below:
# perf c2c report -d peer --coalesce tid,pid,iaddr,dso -N --stdio
[...]
=================================================
Shared Data Cache Line Table
=================================================
#
# ----------- Cacheline ---------- Snoop ------- Load Hitm ------- Snoop Total Total Total --------- Stores -------- ----- Core Load Hit ----- - LLC Load Hit -- - RMT Load Hit -- --- Load Dram ----
# Index Address Node PA cnt Peer Total LclHitm RmtHitm Peer records Loads Stores L1Hit L1Miss N/A FB L1 L2 LclHit LclHitm RmtHit RmtHitm Lcl Rmt
# ..... .................. .... ...... ....... ....... ....... ....... ....... ....... ....... ....... ....... ....... ....... ....... ....... ....... ........ ....... ........ ....... ........ ........
#
0 0xaaaac17d6000 N/A 0 100.00% 0 0 0 99 18851 18851 0 0 0 0 0 18752 0 0 0 0 0 0 0
=================================================
Shared Cache Line Distribution Pareto
=================================================
#
# ----- HITM ----- Snoop ------- Store Refs ------ --------- Data address --------- --------------- cycles --------------- Total cpu Shared
# Num RmtHitm LclHitm Peer L1 Hit L1 Miss N/A Offset Node PA cnt Pid Tid Code address rmt hitm lcl hitm load peer records cnt Symbol Object Source:Line Node{cpus %peers %stores}
# ..... ....... ....... ....... ....... ....... ....... .................. .... ...... ....... ................. .................. ........ ........ ........ ........ ....... ........ ...................... ................ ............... ....
#
-------------------------------------------------------------------------------
0 0 0 99 0 0 0 0xaaaac17d6000
-------------------------------------------------------------------------------
0.00% 0.00% 3.03% 0.00% 0.00% 0.00% 0x20 N/A 0 3603 3603:memstress 0xaaaac17c25ac 0 0 41 376 9314 2 [.] 0x00000000000025ac memstress memstress[25ac] 0{ 2 100.0% n/a}
0.00% 0.00% 3.03% 0.00% 0.00% 0.00% 0x20 N/A 0 3603 3606:memstress 0xaaaac17c25ac 0 0 44 375 9155 1 [.] 0x00000000000025ac memstress memstress[25ac] 0{ 1 100.0% n/a}
0.00% 0.00% 45.45% 0.00% 0.00% 0.00% 0x29 N/A 0 3603 3603:memstress 0xaaaac17c3e88 0 0 175 180 70 2 [.] 0x0000000000003e88 memstress memstress[3e88] 0{ 2 100.0% n/a}
0.00% 0.00% 48.48% 0.00% 0.00% 0.00% 0x29 N/A 0 3603 3606:memstress 0xaaaac17c3e88 0 0 170 180 65 1 [.] 0x0000000000003e88 memstress memstress[3e88] 0{ 1 100.0% n/a}
[...]
Signed-off-by: Leo Yan <[email protected]>
Tested-by: Ali Saidi <[email protected]>
---
tools/perf/builtin-c2c.c | 63 +++++++++++++++++++++++++++++++---------
1 file changed, 49 insertions(+), 14 deletions(-)
diff --git a/tools/perf/builtin-c2c.c b/tools/perf/builtin-c2c.c
index 04f33dc8ca0f..9b6d2a9553e1 100644
--- a/tools/perf/builtin-c2c.c
+++ b/tools/perf/builtin-c2c.c
@@ -117,13 +117,15 @@ enum {
DISPLAY_LCL_HITM,
DISPLAY_RMT_HITM,
DISPLAY_TOT_HITM,
+ DISPLAY_SNP_PEER,
DISPLAY_MAX,
};
static const char *display_str[DISPLAY_MAX] = {
- [DISPLAY_LCL_HITM] = "Local",
- [DISPLAY_RMT_HITM] = "Remote",
- [DISPLAY_TOT_HITM] = "Total",
+ [DISPLAY_LCL_HITM] = "Local HITMs",
+ [DISPLAY_RMT_HITM] = "Remote HITMs",
+ [DISPLAY_TOT_HITM] = "Total HITMs",
+ [DISPLAY_SNP_PEER] = "Snoop Peers",
};
static const struct option c2c_options[] = {
@@ -818,6 +820,11 @@ static double percent_costly_snoop(struct c2c_hist_entry *c2c_he)
case DISPLAY_TOT_HITM:
st = stats->tot_hitm;
tot = total->tot_hitm;
+ break;
+ case DISPLAY_SNP_PEER:
+ st = stats->ld_peer;
+ tot = total->ld_peer;
+ break;
default:
break;
}
@@ -1193,6 +1200,10 @@ node_entry(struct perf_hpp_fmt *fmt __maybe_unused, struct perf_hpp *hpp,
ret = display_metrics(hpp, stats->tot_hitm,
c2c_he->stats.tot_hitm);
break;
+ case DISPLAY_SNP_PEER:
+ ret = display_metrics(hpp, stats->ld_peer,
+ c2c_he->stats.ld_peer);
+ break;
default:
break;
}
@@ -1548,6 +1559,7 @@ static struct c2c_header percent_costly_snoop_header[] = {
[DISPLAY_LCL_HITM] = HEADER_BOTH("Lcl", "Hitm"),
[DISPLAY_RMT_HITM] = HEADER_BOTH("Rmt", "Hitm"),
[DISPLAY_TOT_HITM] = HEADER_BOTH("Tot", "Hitm"),
+ [DISPLAY_SNP_PEER] = HEADER_BOTH("Snoop", "Peer"),
};
static struct c2c_dimension dim_percent_costly_snoop = {
@@ -2024,6 +2036,10 @@ static bool he__display(struct hist_entry *he, struct c2c_stats *stats)
he->filtered = filter_display(c2c_he->stats.tot_hitm,
stats->tot_hitm);
break;
+ case DISPLAY_SNP_PEER:
+ he->filtered = filter_display(c2c_he->stats.ld_peer,
+ stats->ld_peer);
+ break;
default:
break;
}
@@ -2052,6 +2068,8 @@ static inline bool is_valid_hist_entry(struct hist_entry *he)
case DISPLAY_TOT_HITM:
has_record = !!c2c_he->stats.tot_hitm;
break;
+ case DISPLAY_SNP_PEER:
+ has_record = !!c2c_he->stats.ld_peer;
default:
break;
}
@@ -2141,7 +2159,10 @@ static int resort_cl_cb(struct hist_entry *he, void *arg __maybe_unused)
}
static struct c2c_header header_node_0 = HEADER_LOW("Node");
-static struct c2c_header header_node_1 = HEADER_LOW("Node{cpus %hitms %stores}");
+static struct c2c_header header_node_1_hitms_stores =
+ HEADER_LOW("Node{cpus %hitms %stores}");
+static struct c2c_header header_node_1_peers_stores =
+ HEADER_LOW("Node{cpus %peers %stores}");
static struct c2c_header header_node_2 = HEADER_LOW("Node{cpu list}");
static void setup_nodes_header(void)
@@ -2151,7 +2172,10 @@ static void setup_nodes_header(void)
dim_node.header = header_node_0;
break;
case 1:
- dim_node.header = header_node_1;
+ if (c2c.display == DISPLAY_SNP_PEER)
+ dim_node.header = header_node_1_peers_stores;
+ else
+ dim_node.header = header_node_1_hitms_stores;
break;
case 2:
dim_node.header = header_node_2;
@@ -2225,13 +2249,15 @@ static int setup_nodes(struct perf_session *session)
}
#define HAS_HITMS(__h) ((__h)->stats.lcl_hitm || (__h)->stats.rmt_hitm)
+#define HAS_PEER(__h) ((__h)->stats.ld_peer)
static int resort_shared_cl_cb(struct hist_entry *he, void *arg __maybe_unused)
{
struct c2c_hist_entry *c2c_he;
c2c_he = container_of(he, struct c2c_hist_entry, he);
- if (HAS_HITMS(c2c_he)) {
+ if ((c2c.display != DISPLAY_SNP_PEER && HAS_HITMS(c2c_he)) ||
+ (c2c.display == DISPLAY_SNP_PEER && HAS_PEER(c2c_he))) {
c2c.shared_clines++;
c2c_add_stats(&c2c.shared_clines_stats, &c2c_he->stats);
}
@@ -2406,7 +2432,7 @@ static void print_c2c_info(FILE *out, struct perf_session *session)
fprintf(out, "%-36s: %s\n", first ? " Events" : "", evsel__name(evsel));
first = false;
}
- fprintf(out, " Cachelines sort on : %s HITMs\n",
+ fprintf(out, " Cachelines sort on : %s\n",
display_str[c2c.display]);
fprintf(out, " Cacheline data grouping : %s\n", c2c.cl_sort);
}
@@ -2563,7 +2589,7 @@ static int perf_c2c_browser__title(struct hist_browser *browser,
{
scnprintf(bf, size,
"Shared Data Cache Line Table "
- "(%lu entries, sorted on %s HITMs)",
+ "(%lu entries, sorted on %s)",
browser->nr_non_filtered_entries,
display_str[c2c.display]);
return 0;
@@ -2769,6 +2795,8 @@ static int setup_display(const char *str)
c2c.display = DISPLAY_RMT_HITM;
else if (!strcmp(display, "lcl"))
c2c.display = DISPLAY_LCL_HITM;
+ else if (!strcmp(display, "peer"))
+ c2c.display = DISPLAY_SNP_PEER;
else {
pr_err("failed: unknown display type: %s\n", str);
return -1;
@@ -2850,6 +2878,7 @@ static int build_cl_output(char *cl_sort, bool no_source)
static int setup_coalesce(const char *coalesce, bool no_source)
{
const char *c = coalesce ?: coalesce_default;
+ const char *sort_str = NULL;
if (asprintf(&c2c.cl_sort, "offset,%s", c) < 0)
return -ENOMEM;
@@ -2857,12 +2886,16 @@ static int setup_coalesce(const char *coalesce, bool no_source)
if (build_cl_output(c2c.cl_sort, no_source))
return -1;
- if (asprintf(&c2c.cl_resort, "offset,%s",
- c2c.display == DISPLAY_TOT_HITM ?
- "tot_hitm" :
- c2c.display == DISPLAY_RMT_HITM ?
- "rmt_hitm,lcl_hitm" :
- "lcl_hitm,rmt_hitm") < 0)
+ if (c2c.display == DISPLAY_TOT_HITM)
+ sort_str = "tot_hitm";
+ else if (c2c.display == DISPLAY_RMT_HITM)
+ sort_str = "rmt_hitm,lcl_hitm";
+ else if (c2c.display == DISPLAY_LCL_HITM)
+ sort_str = "lcl_hitm,rmt_hitm";
+ else if (c2c.display == DISPLAY_SNP_PEER)
+ sort_str = "ld_peer";
+
+ if (asprintf(&c2c.cl_resort, "offset,%s", sort_str) < 0)
return -ENOMEM;
pr_debug("coalesce sort fields: %s\n", c2c.cl_sort);
@@ -3023,6 +3056,8 @@ static int perf_c2c__report(int argc, const char **argv)
sort_str = "rmt_hitm";
else if (c2c.display == DISPLAY_LCL_HITM)
sort_str = "lcl_hitm";
+ else if (c2c.display == DISPLAY_SNP_PEER)
+ sort_str = "ld_peer";
c2c_hists__reinit(&c2c.hists, output_str, sort_str);
--
2.25.1
Perf c2c tool has an assumption that it heavily depends on HITM snoop
type to detect cache false sharing, unfortunately, HITM is not supported
on some architectures.
Essentially, perf c2c tool wants to find some very costly snooping
operations for false cache sharing, this means it's not necessarily
to stick using HITM tags and we can explore other snooping types
(e.g. SNOOPX_PEER).
For this reason, this patch renames HITM related display macros with
suffix '_HITM', so it can be distinct if later add more display types
for on other snooping type.
Signed-off-by: Leo Yan <[email protected]>
Tested-by: Ali Saidi <[email protected]>
---
tools/perf/builtin-c2c.c | 58 ++++++++++++++++++++--------------------
1 file changed, 29 insertions(+), 29 deletions(-)
diff --git a/tools/perf/builtin-c2c.c b/tools/perf/builtin-c2c.c
index baa003c150b1..180208a4085f 100644
--- a/tools/perf/builtin-c2c.c
+++ b/tools/perf/builtin-c2c.c
@@ -114,16 +114,16 @@ struct perf_c2c {
};
enum {
- DISPLAY_LCL,
- DISPLAY_RMT,
- DISPLAY_TOT,
+ DISPLAY_LCL_HITM,
+ DISPLAY_RMT_HITM,
+ DISPLAY_TOT_HITM,
DISPLAY_MAX,
};
static const char *display_str[DISPLAY_MAX] = {
- [DISPLAY_LCL] = "Local",
- [DISPLAY_RMT] = "Remote",
- [DISPLAY_TOT] = "Total",
+ [DISPLAY_LCL_HITM] = "Local",
+ [DISPLAY_RMT_HITM] = "Remote",
+ [DISPLAY_TOT_HITM] = "Total",
};
static const struct option c2c_options[] = {
@@ -807,15 +807,15 @@ static double percent_hitm(struct c2c_hist_entry *c2c_he)
total = &hists->stats;
switch (c2c.display) {
- case DISPLAY_RMT:
+ case DISPLAY_RMT_HITM:
st = stats->rmt_hitm;
tot = total->rmt_hitm;
break;
- case DISPLAY_LCL:
+ case DISPLAY_LCL_HITM:
st = stats->lcl_hitm;
tot = total->lcl_hitm;
break;
- case DISPLAY_TOT:
+ case DISPLAY_TOT_HITM:
st = stats->tot_hitm;
tot = total->tot_hitm;
default:
@@ -1181,15 +1181,15 @@ node_entry(struct perf_hpp_fmt *fmt __maybe_unused, struct perf_hpp *hpp,
advance_hpp(hpp, ret);
switch (c2c.display) {
- case DISPLAY_RMT:
+ case DISPLAY_RMT_HITM:
ret = display_metrics(hpp, stats->rmt_hitm,
c2c_he->stats.rmt_hitm);
break;
- case DISPLAY_LCL:
+ case DISPLAY_LCL_HITM:
ret = display_metrics(hpp, stats->lcl_hitm,
c2c_he->stats.lcl_hitm);
break;
- case DISPLAY_TOT:
+ case DISPLAY_TOT_HITM:
ret = display_metrics(hpp, stats->tot_hitm,
c2c_he->stats.tot_hitm);
break;
@@ -1545,9 +1545,9 @@ static struct c2c_dimension dim_tot_loads = {
};
static struct c2c_header percent_hitm_header[] = {
- [DISPLAY_LCL] = HEADER_BOTH("Lcl", "Hitm"),
- [DISPLAY_RMT] = HEADER_BOTH("Rmt", "Hitm"),
- [DISPLAY_TOT] = HEADER_BOTH("Tot", "Hitm"),
+ [DISPLAY_LCL_HITM] = HEADER_BOTH("Lcl", "Hitm"),
+ [DISPLAY_RMT_HITM] = HEADER_BOTH("Rmt", "Hitm"),
+ [DISPLAY_TOT_HITM] = HEADER_BOTH("Tot", "Hitm"),
};
static struct c2c_dimension dim_percent_hitm = {
@@ -2018,15 +2018,15 @@ static bool he__display(struct hist_entry *he, struct c2c_stats *stats)
c2c_he = container_of(he, struct c2c_hist_entry, he);
switch (c2c.display) {
- case DISPLAY_LCL:
+ case DISPLAY_LCL_HITM:
he->filtered = filter_display(c2c_he->stats.lcl_hitm,
stats->lcl_hitm);
break;
- case DISPLAY_RMT:
+ case DISPLAY_RMT_HITM:
he->filtered = filter_display(c2c_he->stats.rmt_hitm,
stats->rmt_hitm);
break;
- case DISPLAY_TOT:
+ case DISPLAY_TOT_HITM:
he->filtered = filter_display(c2c_he->stats.tot_hitm,
stats->tot_hitm);
break;
@@ -2049,13 +2049,13 @@ static inline bool is_valid_hist_entry(struct hist_entry *he)
return true;
switch (c2c.display) {
- case DISPLAY_LCL:
+ case DISPLAY_LCL_HITM:
has_record = !!c2c_he->stats.lcl_hitm;
break;
- case DISPLAY_RMT:
+ case DISPLAY_RMT_HITM:
has_record = !!c2c_he->stats.rmt_hitm;
break;
- case DISPLAY_TOT:
+ case DISPLAY_TOT_HITM:
has_record = !!c2c_he->stats.tot_hitm;
break;
default:
@@ -2752,11 +2752,11 @@ static int setup_display(const char *str)
const char *display = str ?: "tot";
if (!strcmp(display, "tot"))
- c2c.display = DISPLAY_TOT;
+ c2c.display = DISPLAY_TOT_HITM;
else if (!strcmp(display, "rmt"))
- c2c.display = DISPLAY_RMT;
+ c2c.display = DISPLAY_RMT_HITM;
else if (!strcmp(display, "lcl"))
- c2c.display = DISPLAY_LCL;
+ c2c.display = DISPLAY_LCL_HITM;
else {
pr_err("failed: unknown display type: %s\n", str);
return -1;
@@ -2846,9 +2846,9 @@ static int setup_coalesce(const char *coalesce, bool no_source)
return -1;
if (asprintf(&c2c.cl_resort, "offset,%s",
- c2c.display == DISPLAY_TOT ?
+ c2c.display == DISPLAY_TOT_HITM ?
"tot_hitm" :
- c2c.display == DISPLAY_RMT ?
+ c2c.display == DISPLAY_RMT_HITM ?
"rmt_hitm,lcl_hitm" :
"lcl_hitm,rmt_hitm") < 0)
return -ENOMEM;
@@ -3005,11 +3005,11 @@ static int perf_c2c__report(int argc, const char **argv)
"ld_rmthit,rmt_hitm,"
"dram_lcl,dram_rmt";
- if (c2c.display == DISPLAY_TOT)
+ if (c2c.display == DISPLAY_TOT_HITM)
sort_str = "tot_hitm";
- else if (c2c.display == DISPLAY_RMT)
+ else if (c2c.display == DISPLAY_RMT_HITM)
sort_str = "rmt_hitm";
- else if (c2c.display == DISPLAY_LCL)
+ else if (c2c.display == DISPLAY_LCL_HITM)
sort_str = "lcl_hitm";
c2c_hists__reinit(&c2c.hists, output_str, sort_str);
--
2.25.1
The node header array contains 3 items, each item is used for one of
the 3 flavors for node accessing info. To extend sorting on other
snooping type and not always stick to HITMs, the second header string
"Node{cpus %hitms %stores}" should be adjusted (e.g. it's changed as
"Node{cpus %peer %stores}").
For this reason, this patch changes the node header array to three
flat variables and uses switch-case in function setup_nodes_header(),
thus it is easier for altering the header string.
Signed-off-by: Leo Yan <[email protected]>
Tested-by: Ali Saidi <[email protected]>
---
tools/perf/builtin-c2c.c | 26 +++++++++++++++++++-------
1 file changed, 19 insertions(+), 7 deletions(-)
diff --git a/tools/perf/builtin-c2c.c b/tools/perf/builtin-c2c.c
index 1e1565e5d396..04f33dc8ca0f 100644
--- a/tools/perf/builtin-c2c.c
+++ b/tools/perf/builtin-c2c.c
@@ -1653,12 +1653,6 @@ static struct c2c_dimension dim_dso = {
.se = &sort_dso,
};
-static struct c2c_header header_node[3] = {
- HEADER_LOW("Node"),
- HEADER_LOW("Node{cpus %hitms %stores}"),
- HEADER_LOW("Node{cpu list}"),
-};
-
static struct c2c_dimension dim_node = {
.name = "node",
.cmp = empty_cmp,
@@ -2146,9 +2140,27 @@ static int resort_cl_cb(struct hist_entry *he, void *arg __maybe_unused)
return 0;
}
+static struct c2c_header header_node_0 = HEADER_LOW("Node");
+static struct c2c_header header_node_1 = HEADER_LOW("Node{cpus %hitms %stores}");
+static struct c2c_header header_node_2 = HEADER_LOW("Node{cpu list}");
+
static void setup_nodes_header(void)
{
- dim_node.header = header_node[c2c.node_info];
+ switch (c2c.node_info) {
+ case 0:
+ dim_node.header = header_node_0;
+ break;
+ case 1:
+ dim_node.header = header_node_1;
+ break;
+ case 2:
+ dim_node.header = header_node_2;
+ break;
+ default:
+ break;
+ }
+
+ return;
}
static int setup_nodes(struct perf_session *session)
--
2.25.1
Since the new display option 'peer' is introduced, this patch is to
update the documentation to reflect it.
Signed-off-by: Leo Yan <[email protected]>
Tested-by: Ali Saidi <[email protected]>
---
tools/perf/Documentation/perf-c2c.txt | 22 +++++++++++++++++-----
1 file changed, 17 insertions(+), 5 deletions(-)
diff --git a/tools/perf/Documentation/perf-c2c.txt b/tools/perf/Documentation/perf-c2c.txt
index 6f69173731aa..cc98a7e774f6 100644
--- a/tools/perf/Documentation/perf-c2c.txt
+++ b/tools/perf/Documentation/perf-c2c.txt
@@ -109,7 +109,8 @@ REPORT OPTIONS
-d::
--display::
- Switch to HITM type (rmt, lcl) to display and sort on. Total HITMs as default.
+ Switch to HITM type (rmt, lcl) or peer snooping type (peer) to display
+ and sort on. Total HITMs (tot) as default.
--stitch-lbr::
Show callgraph with stitched LBRs, which may have more complete
@@ -174,12 +175,18 @@ For each cacheline in the 1) list we display following data:
Cacheline
- cacheline address (hex number)
- Rmt/Lcl Hitm
+ Rmt/Lcl Hitm (For display with HITM types)
- cacheline percentage of all Remote/Local HITM accesses
+ Snoop Peer (For display with peer type)
+ - cacheline percentage of peer access
+
LLC Load Hitm - Total, LclHitm, RmtHitm
- count of Total/Local/Remote load HITMs
+ Snoop Peer
+ - count of peer access
+
Total records
- sum of all cachelines accesses
@@ -211,6 +218,9 @@ For each offset in the 2) list we display following data:
HITM - Rmt, Lcl
- % of Remote/Local HITM accesses for given offset within cacheline
+ Snoop Peer
+ - % of peer accesses for given offset within cacheline
+
Store Refs - L1 Hit, L1 Miss, N/A
- % of store accesses that hit L1, missed L1 and N/A (no available) memory
level for given offset within cacheline
@@ -227,8 +237,9 @@ For each offset in the 2) list we display following data:
Code address
- code address responsible for the accesses
- cycles - rmt hitm, lcl hitm, load
- - sum of cycles for given accesses - Remote/Local HITM and generic load
+ cycles - rmt hitm, lcl hitm, load, peer
+ - sum of cycles for given accesses - Remote/Local HITM, generic load and
+ peer access
cpu cnt
- number of cpus that participated on the access
@@ -251,7 +262,8 @@ The 'Node' field displays nodes that accesses given cacheline
offset. Its output comes in 3 flavors:
- node IDs separated by ','
- node IDs with stats for each ID, in following format:
- Node{cpus %hitms %stores}
+ Node{cpus %hitms %stores} (For display with HITM types)
+ Node{cpus %peers %stores} (For display with "peer" type)
- node IDs with list of affected CPUs in following format:
Node{cpu list}
--
2.25.1
On 5/8/2022 5:23 PM, Leo Yan wrote:
> Except the existed three display options 'tot', 'rmt', 'lcl', this patch
> adds a new option 'peer' so can sort on the cache hit for peer snooping.
>
> For displaying with option 'peer', the "Shared Data Cache Line Table" and
> "Shared Cache Line Distribution Pareto" both sort with the metrics
> "ld_peer". As result, we can get the 'peer' display as below:
>
> # perf c2c report -d peer --coalesce tid,pid,iaddr,dso -N --stdio
>
Hi Leo,
I tested v2 patch on 2P Altra system.
In case the false-sharing data is mainly from remote node, 'Snoop Peers'
cannot indicate severity of false-sharing. As showed in bellow output,
there are only 10 'Load HIT Peer' records, while there are 2353
'Load Remote DRAM' records.
And the name 'Load Remote DRAM' is kind of misleading, since we cannot tell
the data source is 'DRAM'.
Run false_sharing test(https://github.com/joemario/perf-c2c-usage-files):
one lock_th on node 0, one reader_thd on node 1:
# perf c2c record -- numactl -m 0 ./false_sharing.exe 1
131 mticks, reader_thd (thread 1), on node 1 (cpu 80).
145 mticks, lock_th (thread 0), on node 0 (cpu 9).
[ perf record: Woken up 16 times to write data ]
[ perf record: Captured and wrote 33.726 MB perf.data ]
# perf c2c report -d peer --coalesce tid,pid,iaddr,dso -N --stdio
Warning:
Arm SPE CONTEXT packets not found in the traces.
Matching of TIDs to SPE events could be inaccurate.
Warning:
AUX data detected collision 6 times out of 47!
=================================================
Trace Event Information
=================================================
Total records : 551944
Locked Load/Store Operations : 0
Load Operations : 493082
Loads - uncacheable : 0
Loads - IO : 0
Loads - Miss : 0
Loads - no mapping : 0
Load Fill Buffer Hit : 0
Load L1D hit : 490589
Load L2D hit : 117
Load LLC hit : 11
Load HIT Peer : 10
Load Local HITM : 0
Load Remote HITM : 0
Load Remote HIT : 0
Load Local DRAM : 2
Load Remote DRAM : 2353
Load MESI State Exclusive : 2355
Load MESI State Shared : 0
Load LLC Misses : 2355
Load access blocked by data : 0
Load access blocked by address : 0
LLC Misses to Local DRAM : 0.1%
LLC Misses to Remote DRAM : 99.9%
LLC Misses to Remote cache (HIT) : 0.0%
LLC Misses to Remote cache (HITM) : 0.0%
Store Operations : 58862
Store - uncacheable : 0
Store - no mapping : 0
Store L1D Hit : 0
Store L1D Miss : 0
Store No available memory level : 58862
No Page Map Rejects : 490
Unable to parse data source : 0
=================================================
Global Shared Cache Line Event Information
=================================================
Total Shared Cache Lines : 9
Load HITs on shared lines : 21
Fill Buffer Hits on shared lines : 0
L1D hits on shared lines : 6
L2D hits on shared lines : 1
Load HITs on peer cache lines : 10
LLC hits on shared lines : 0
Locked Access on shared lines : 0
Blocked Access on shared lines : 0
Store HITs on shared lines : 0
Store L1D hits on shared lines : 0
Store No available memory level : 0
Total Merged records : 0
=================================================
c2c details
=================================================
Events : arm_spe_0/ts_enable=1,load_filter=1,store_filter=1,min_latency=30/
: dummy:u
: memory
Cachelines sort on : Snoop Peers
Cacheline data grouping : offset,tid,pid,iaddr,dso
[...]
Thanks,
-adam
Hi Adam,
On Fri, May 13, 2022 at 05:05:45PM +0800, Adam Li wrote:
[...]
> > As result, we can get the 'peer' display as below:
> >
> > # perf c2c report -d peer --coalesce tid,pid,iaddr,dso -N --stdio
> >
>
> Hi Leo,
>
> I tested v2 patch on 2P Altra system.
> In case the false-sharing data is mainly from remote node, 'Snoop Peers'
> cannot indicate severity of false-sharing. As showed in bellow output,
> there are only 10 'Load HIT Peer' records, while there are 2353
> 'Load Remote DRAM' records.
>
> And the name 'Load Remote DRAM' is kind of misleading, since we cannot tell
> the data source is 'DRAM'.
Appreicate for the testing, and your feedback is very helpful.
After some discussion, in the patch set v9 "perf: arm-spe: Decode SPE
source and use for perf c2c" (please ignore the typo "v8", actually
it's patch set v9), Ali has updated the memory tag setting for the
type ARM_SPE_NV_REMOTE. I think the prominent change is we set
PERF_MEM_SNOOPX_PEER flag for remote access.
Please note, in the total statistics, all remote accesses will be
accounted into metric "rmt_hit", so "rmt_hit" includes the access for
remote DRAM or any upwards cache levels due we cannot distinguish
them.
From my experiment, with this updating the output result is promised
for the peer accesses and it's easier for inspecting false sharing.
As you might see I have prepared a git repo:
https://git.linaro.org/people/leo.yan/linux-spe.git/ branch:
perf_c2c_arm_spe_peer_v3, which contains the updated patches for both
memory flag setting and perf c2c related patches.
Could you confirm if the updated code works for you or not?
Thanks,
Leo
Hi Leo,
Thanks for the update.
On 5/18/2022 2:12 PM, Leo Yan wrote:
> Please note, in the total statistics, all remote accesses will be
> accounted into metric "rmt_hit", so "rmt_hit" includes the access for
> remote DRAM or any upwards cache levels due we cannot distinguish
> them.
>
Agree that "Load Remote HIT" makes more sense than "Load Remote DRAM".
> From my experiment, with this updating the output result is promised
> for the peer accesses and it's easier for inspecting false sharing.
>
> As you might see I have prepared a git repo:
> https://git.linaro.org/people/leo.yan/linux-spe.git/ branch:
> perf_c2c_arm_spe_peer_v3, which contains the updated patches for both
> memory flag setting and perf c2c related patches.
>
> Could you confirm if the updated code works for you or not?
>
I tested v3 patch (perf_c2c_arm_spe_peer_v3 branch) on 2P Altra system.
Compared with v2, "Snoop Peer" can better indicate cache false-sharing,
for the 'false_sharing.exe' test case.
Bellow are details:
# perf c2c record -- numactl -m 0 ./false_sharing.exe 2
183 mticks, reader_thd (thread 2), on node 0 (cpu 78).
195 mticks, reader_thd (thread 3), on node 1 (cpu 124).
546 mticks, lock_th (thread 0), on node 0 (cpu 0).
562 mticks, lock_th (thread 1), on node 1 (cpu 123).
[ perf record: Woken up 36 times to write data ]
[ perf record: Captured and wrote 72.440 MB perf.data ]
# perf c2c report -d peer --coalesce tid,pid,iaddr,dso -N --stdio
Warning:
Arm SPE CONTEXT packets not found in the traces.
Matching of TIDs to SPE events could be inaccurate.
Warning:
AUX data detected collision 20 times out of 168!
=================================================
Total records : 1198728
Locked Load/Store Operations : 0
Load Operations : 1031196
Loads - uncacheable : 0
Loads - IO : 0
Loads - Miss : 0
Loads - no mapping : 0
Load Fill Buffer Hit : 0
Load L1D hit : 970636
Load L2D hit : 292
Load LLC hit : 2477
Load Local HITM : 0
Load Remote HITM : 0
Load Remote HIT : 56459
Load Local DRAM : 1332
Load Remote DRAM : 0
Load MESI State Exclusive : 1332
Load MESI State Shared : 0
Load LLC Misses : 57791
Load access blocked by data : 0
Load access blocked by address : 0
Load HIT Peer : 58814
LLC Misses to Local DRAM : 2.3%
LLC Misses to Remote DRAM : 0.0%
LLC Misses to Remote cache (HIT) : 97.7%
LLC Misses to Remote cache (HITM) : 0.0%
Store Operations : 167532
Store - uncacheable : 0
Store - no mapping : 0
Store L1D Hit : 0
Store L1D Miss : 0
Store No available memory level : 167532
No Page Map Rejects : 1234
Unable to parse data source : 0
=================================================
Global Shared Cache Line Event Information
=================================================
Total Shared Cache Lines : 45
Load HITs on shared lines : 226254
Fill Buffer Hits on shared lines : 0
L1D hits on shared lines : 166010
L2D hits on shared lines : 4
Load HITs on peer cache lines : 58814
LLC hits on shared lines : 2455
Locked Access on shared lines : 0
Blocked Access on shared lines : 0
Store HITs on shared lines : 96403
Store L1D hits on shared lines : 0
Store No available memory level : 96403
Total Merged records : 96403
=================================================
c2c details
=================================================
Events : arm_spe_0/ts_enable=1,load_filter=1,store_filter=1,min_latency=30/
: dummy:u
: memory
Cachelines sort on : Snoop Peers
Cacheline data grouping : offset,tid,pid,iaddr,dso
=================================================
Shared Data Cache Line Table
=================================================
#
# ----------- Cacheline ---------- Snoop ------- Load Hitm ------- Snoop Total Total Total --------- Stores -------- ----- Core Load Hit ----- - LLC Load Hit -- - RMT Load Hit -- --- Load Dram ----
# Index Address Node PA cnt Peer Total LclHitm RmtHitm Peer records Loads Stores L1Hit L1Miss N/A FB L1 L2 LclHit LclHitm RmtHit RmtHitm Lcl Rmt
# ..... .................. .... ...... ....... ....... ....... ....... ....... ....... ....... ....... ....... ....... ....... ....... ....... ....... ........ ....... ........ ....... ........ ........
#
0 0x420180 N/A 0 95.53% 0 0 0 56183 246056 219522 26534 0 0 26534 0 161914 0 106 0 56176 0 1326 0
1 0x420100 N/A 0 4.37% 0 0 0 2571 76437 6576 69861 0 0 69861 0 4005 0 2335 0 236 0 0 0
[...]
Thanks,
-adam
On 08/05/2022 10:23, Leo Yan wrote:
> Arm64 Neoverse CPUs supports data source in Arm SPE trace, this allows
> us to detect cache line contention and transfers.
>
> Unlike x86 architecture, Arm SPE trace data cannot provide 'HITM'
> snooping flag, Ali Said has a patch set "perf: arm-spe: Decode SPE
> source and use for perf c2c" [1] which introduces 'peer' flag and
> synthesizes memory samples with this flag.
>
> Based on patch set [1], this patch set is to finish the second half work
> to consume the 'peer' flag in perf c2c tool, it adds an extra display
> 'peer' mode.
>
> Patches 01, 02 and 03 are to support 'N/A' metrics for store operations.
>
> Patches 04 and 05 adds statistics and dimensions for memory samples with
> peer flag.
>
> Patches 06, 07, 08 are for refactoring, it refines the code with more
> general naming so this can allow us to easier to extend display modes
> but not strictly bound to HITM tags.
>
> Patches 09, 10 and 11 are to extend display 'peer' mode, it also updates
> the document and also changes to use 'peer' mode as default mode on
> Arm64 arches.
>
> This patch set has been verified for both x86 and Arm64 memory samples.
>
> The display result with x86 memory samples:
>
> =================================================
> Shared Data Cache Line Table
> =================================================
> #
> # ----------- Cacheline ---------- Tot ------- Load Hitm ------- Snoop Total Total Total --------- Stores -------- ----- Core Load Hit ----- - LLC Load Hit -- - RMT Load Hit -- --- Load Dram ----
> # Index Address Node PA cnt Hitm Total LclHitm RmtHitm Peer records Loads Stores L1Hit L1Miss N/A FB L1 L2 LclHit LclHitm RmtHit RmtHitm Lcl Rmt
> # ..... .................. .... ...... ....... ....... ....... ....... ....... ....... ....... ....... ....... ....... ....... ....... ....... ....... ........ ....... ........ ....... ........ ........
> #
> 0 0x55c8971f0080 0 1967 66.14% 252 252 0 0 6044 3550 2494 2024 470 0 528 2672 78 20 252 0 0 0 0
> 1 0x55c8971f00c0 0 1 33.86% 129 129 0 0 914 914 0 0 0 0 272 374 52 87 129 0 0 0 0
>
> =================================================
> Shared Cache Line Distribution Pareto
> =================================================
> #
> # ----- HITM ----- Snoop ------- Store Refs ------ --------- Data address --------- --------------- cycles --------------- Total cpu Shared
> # Num RmtHitm LclHitm Peer L1 Hit L1 Miss N/A Offset Node PA cnt Code address rmt hitm lcl hitm load peer records cnt Symbol Object Source:Line Node
> # ..... ....... ....... ....... ....... ....... ....... .................. .... ...... .................. ........ ........ ........ ........ ....... ........ ...................... ................. ....................... ....
> #
> -------------------------------------------------------------------------------
> 0 0 252 0 2024 470 0 0x55c8971f0080
> -------------------------------------------------------------------------------
> 0.00% 12.30% 0.00% 0.00% 0.00% 0.00% 0x0 0 1 0x55c8971ed3e9 0 1313 863 0 1222 3 [.] 0x00000000000013e9 false_sharing.exe false_sharing.exe[13e9] 0
> 0.00% 0.79% 0.00% 90.51% 0.00% 0.00% 0x0 0 1 0x55c8971ed3e2 0 1800 878 0 3029 3 [.] 0x00000000000013e2 false_sharing.exe false_sharing.exe[13e2] 0
> 0.00% 0.00% 0.00% 9.49% 100.00% 0.00% 0x0 0 1 0x55c8971ed3f4 0 0 0 0 662 3 [.] 0x00000000000013f4 false_sharing.exe false_sharing.exe[13f4] 0
> 0.00% 86.90% 0.00% 0.00% 0.00% 0.00% 0x20 0 1 0x55c8971ed447 0 141 103 0 1131 2 [.] 0x0000000000001447 false_sharing.exe false_sharing.exe[1447] 0
>
> -------------------------------------------------------------------------------
> 1 0 129 0 0 0 0 0x55c8971f00c0
> -------------------------------------------------------------------------------
> 0.00% 100.00% 0.00% 0.00% 0.00% 0.00% 0x20 0 1 0x55c8971ed455 0 88 94 0 914 2 [.] 0x0000000000001455 false_sharing.exe false_sharing.exe[1455] 0
>
>
> The display result with Arm SPE memory samples:
>
> =================================================
> Shared Data Cache Line Table
> =================================================
> #
> # ----------- Cacheline ---------- Snoop ------- Load Hitm ------- Snoop Total Total Total --------- Stores -------- ----- Core Load Hit ----- - LLC Load Hit -- - RMT Load Hit -- --- Load Dram ----
> # Index Address Node PA cnt Peer Total LclHitm RmtHitm Peer records Loads Stores L1Hit L1Miss N/A FB L1 L2 LclHit LclHitm RmtHit RmtHitm Lcl Rmt
> # ..... .................. .... ...... ....... ....... ....... ....... ....... ....... ....... ....... ....... ....... ....... ....... ....... ....... ........ ....... ........ ....... ........ ........
> #
> 0 0xaaaac17d6000 N/A 0 100.00% 0 0 0 99 18851 18851 0 0 0 0 0 18752 0 0 0 0 0 0 0
>
> =================================================
> Shared Cache Line Distribution Pareto
> =================================================
> #
> # ----- HITM ----- Snoop ------- Store Refs ------ --------- Data address --------- --------------- cycles --------------- Total cpu Shared
> # Num RmtHitm LclHitm Peer L1 Hit L1 Miss N/A Offset Node PA cnt Code address rmt hitm lcl hitm load peer records cnt Symbol Object Source:Line Node
> # ..... ....... ....... ....... ....... ....... ....... .................. .... ...... .................. ........ ........ ........ ........ ....... ........ ...................... ................ ............... ....
> #
> -------------------------------------------------------------------------------
> 0 0 0 99 0 0 0 0xaaaac17d6000
> -------------------------------------------------------------------------------
> 0.00% 0.00% 6.06% 0.00% 0.00% 0.00% 0x20 N/A 0 0xaaaac17c25ac 0 0 43 375 18469 2 [.] 0x00000000000025ac memstress memstress[25ac] 0
> 0.00% 0.00% 93.94% 0.00% 0.00% 0.00% 0x29 N/A 0 0xaaaac17c3e88 0 0 173 180 135 2 [.] 0x0000000000003e88 memstress memstress[3e88] 0
>
Hi Leo,
I see that the HITM and Store Refs columns (even N/A) are all 0% on Arm.
Assuming it's a similar example to the x86 one above, does that mean that
these are always expected to be 0? If so, should they be removed to avoid
confusion, or is it just a case of the example not producing any non 0
data for these columns?
Thanks
James
> [1] https://lore.kernel.org/lkml/[email protected]/
>
> Changes from v1:
> * Update patches 01, 02 and 03 to support 'N/A' metrics for store
> operations, so can align with the patch set [1] for store samples.
>
>
> Leo Yan (11):
> perf mem: Add stats for store operation with no available memory level
> perf c2c: Add dimensions for 'N/A' metrics of store operation
> perf c2c: Update documentation for store metric 'N/A'
> perf mem: Add statistics for peer snooping
> perf c2c: Add dimensions for peer load operations
> perf c2c: Use explicit names for display macros
> perf c2c: Rename dimension from 'percent_hitm' to
> 'percent_costly_snoop'
> perf c2c: Refactor node header
> perf c2c: Sort on peer snooping for load operations
> perf c2c: Update documentation for new display option 'peer'
> perf c2c: Use 'peer' as default display for Arm64
>
> tools/perf/Documentation/perf-c2c.txt | 30 ++-
> tools/perf/builtin-c2c.c | 363 ++++++++++++++++++++------
> tools/perf/util/mem-events.c | 14 +-
> tools/perf/util/mem-events.h | 2 +
> 4 files changed, 323 insertions(+), 86 deletions(-)
>
On Thu, May 19, 2022 at 05:06:18PM +0800, Adam Li wrote:
> Hi Leo,
>
> Thanks for the update.
> On 5/18/2022 2:12 PM, Leo Yan wrote:
>
> > Please note, in the total statistics, all remote accesses will be
> > accounted into metric "rmt_hit", so "rmt_hit" includes the access for
> > remote DRAM or any upwards cache levels due we cannot distinguish
> > them.
> >
>
> Agree that "Load Remote HIT" makes more sense than "Load Remote DRAM".
>
> > From my experiment, with this updating the output result is promised
> > for the peer accesses and it's easier for inspecting false sharing.
> >
> > As you might see I have prepared a git repo:
> > https://git.linaro.org/people/leo.yan/linux-spe.git/ branch:
> > perf_c2c_arm_spe_peer_v3, which contains the updated patches for both
> > memory flag setting and perf c2c related patches.
> >
> > Could you confirm if the updated code works for you or not?
> >
>
> I tested v3 patch (perf_c2c_arm_spe_peer_v3 branch) on 2P Altra system.
>
> Compared with v2, "Snoop Peer" can better indicate cache false-sharing,
> for the 'false_sharing.exe' test case.
Thanks a lot for the testing and confirmation, Adam.
I will update the patches based on the comments from Joe and James,
after get ready will send out the patch set and loop you.
Thanks,
Leo
Hi James,
On Thu, May 19, 2022 at 03:19:09PM +0100, James Clark wrote:
[...]
> Hi Leo,
>
> I see that the HITM and Store Refs columns (even N/A) are all 0% on Arm.
> Assuming it's a similar example to the x86 one above, does that mean that
> these are always expected to be 0? If so, should they be removed to avoid
> confusion, or is it just a case of the example not producing any non 0
> data for these columns?
On Arm SPE, we need to use 'N/A' for Store Refs, but we doesn't need
HITM.
On x86, we should expect 'N/A' is useless and should be removed.
I will update the view for this.
Thanks for the testing!
Leo