ligang/vpp
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

perfmon plugin: 2-way parallel stat collection

Browse Source 
As a FUD reduction measure, this patch implements 2-way parallel
counter collection. Synthetic stat component counter pairs run at the
same time. Running two counters (of any kind) at the same time
naturally reduces the aggregate time required by an approximate
factor-of-2, depending on whether an even or odd number of stats have
been requested.

I don't completely buy the argument that computing synthetic stats
such as instructions-per-clock will be inaccurate if component counter
values are collected sequentially. Given uniform traffic pattern, it
must make no difference.

As the collection interval increases, the difference between serial
and parallel component counter collection will approach zero, see also
the Central Limit theorem.

Change-Id: I36ebdcf125e8882cca8a1929ec58f17fba1ad8f1
Signed-off-by: Dave Barach <dave@barachs.net>
This commit is contained in:
Dave Barach
2019-01-24 10:34:24 -05:00
committed by Damjan Marion
parent 22f23ae802
commit ec595ef026
8 changed files with 289 additions and 215 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/plugins/perfmon/perfmon.c
+41 -21
View File
@@ -157,10 +157,16 @@ perfmon_init (vlib_main_t * vm)
pm->log_class = vlib_log_register_class ("perfmon", 0);
/* Default data collection interval */
pm->timeout_interval = 3.0;
vec_validate (pm->pm_fds, vec_len (vlib_mains) - 1);
vec_validate (pm->perf_event_pages, vec_len (vlib_mains) - 1);
vec_validate (pm->rdpmc_indices, vec_len (vlib_mains) - 1);
pm->timeout_interval = 2.0; /* seconds */
vec_validate (pm->pm_fds, 1);
vec_validate (pm->pm_fds[0], vec_len (vlib_mains) - 1);
vec_validate (pm->pm_fds[1], vec_len (vlib_mains) - 1);
vec_validate (pm->perf_event_pages, 1);
vec_validate (pm->perf_event_pages[0], vec_len (vlib_mains) - 1);
vec_validate (pm->perf_event_pages[1], vec_len (vlib_mains) - 1);
vec_validate (pm->rdpmc_indices, 1);
vec_validate (pm->rdpmc_indices[0], vec_len (vlib_mains) - 1);
vec_validate (pm->rdpmc_indices[1], vec_len (vlib_mains) - 1);
pm->page_size = getpagesize ();
ht = pm->perfmon_table = 0;
@@ -297,10 +303,12 @@ set_pmc_command_fn (vlib_main_t * vm,
perfmon_main_t *pm = &perfmon_main;
unformat_input_t _line_input, *line_input = &_line_input;
perfmon_event_config_t ec;
f64 delay;
u32 timeout_seconds;
u32 deadman;
vec_reset_length (pm->events_to_collect);
vec_reset_length (pm->single_events_to_collect);
vec_reset_length (pm->paired_events_to_collect);
pm->ipc_event_index = ~0;
pm->mispredict_event_index = ~0;
@@ -316,28 +324,28 @@ set_pmc_command_fn (vlib_main_t * vm,
ec.name = "instructions";
ec.pe_type = PERF_TYPE_HARDWARE;
ec.pe_config = PERF_COUNT_HW_INSTRUCTIONS;
pm->ipc_event_index = vec_len (pm->events_to_collect);
vec_add1 (pm->events_to_collect, ec);
pm->ipc_event_index = vec_len (pm->paired_events_to_collect);
vec_add1 (pm->paired_events_to_collect, ec);
ec.name = "cpu-cycles";
ec.pe_type = PERF_TYPE_HARDWARE;
ec.pe_config = PERF_COUNT_HW_CPU_CYCLES;
vec_add1 (pm->events_to_collect, ec);
vec_add1 (pm->paired_events_to_collect, ec);
}
else if (unformat (line_input, "branch-mispredict-rate"))
{
ec.name = "branch-misses";
ec.pe_type = PERF_TYPE_HARDWARE;
ec.pe_config = PERF_COUNT_HW_BRANCH_MISSES;
pm->mispredict_event_index = vec_len (pm->events_to_collect);
vec_add1 (pm->events_to_collect, ec);
pm->mispredict_event_index = vec_len (pm->paired_events_to_collect);
vec_add1 (pm->paired_events_to_collect, ec);
ec.name = "branches";
ec.pe_type = PERF_TYPE_HARDWARE;
ec.pe_config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS;
vec_add1 (pm->events_to_collect, ec);
vec_add1 (pm->paired_events_to_collect, ec);
}
else if (unformat (line_input, "%U", unformat_processor_event, pm, &ec))
{
vec_add1 (pm->events_to_collect, ec);
vec_add1 (pm->single_events_to_collect, ec);
}
#define _(type,event,str) \
else if (unformat (line_input, str)) \
@@ -345,7 +353,7 @@ set_pmc_command_fn (vlib_main_t * vm,
ec.name = str; \
ec.pe_type = type; \
ec.pe_config = event; \
vec_add1 (pm->events_to_collect, ec); \
vec_add1 (pm->single_events_to_collect, ec); \
}
foreach_perfmon_event
#undef _
@@ -354,21 +362,33 @@ set_pmc_command_fn (vlib_main_t * vm,
format_unformat_error, line_input);
}
if (vec_len (pm->events_to_collect) == 0)
/* Stick paired events at the front of the (unified) list */
if (vec_len (pm->paired_events_to_collect) > 0)
{
perfmon_event_config_t *tmp;
/* first 2n events are pairs... */
vec_append (pm->paired_events_to_collect, pm->single_events_to_collect);
tmp = pm->single_events_to_collect;
pm->single_events_to_collect = pm->paired_events_to_collect;
pm->paired_events_to_collect = tmp;
}
if (vec_len (pm->single_events_to_collect) == 0)
return clib_error_return (0, "no events specified...");
/* Figure out how long data collection will take */
delay =
((f64) vec_len (pm->single_events_to_collect)) * pm->timeout_interval;
delay /= 2.0; /* collect 2 stats at once */
vlib_cli_output (vm, "Start collection for %d events, wait %.2f seconds",
vec_len (pm->events_to_collect),
(f64) (vec_len (pm->events_to_collect))
* pm->timeout_interval);
vec_len (pm->single_events_to_collect), delay);
vlib_process_signal_event (pm->vlib_main, perfmon_periodic_node.index,
PERFMON_START, 0);
/* Coarse-grained wait */
vlib_process_suspend (vm,
((f64) (vec_len (pm->events_to_collect)
* pm->timeout_interval)));
vlib_process_suspend (vm, delay);
deadman = 0;
/* Reasonable to guess that collection may not be quite done... */
@@ -438,7 +458,7 @@ format_capture (u8 * s, va_list * args)
if (i == pm->ipc_event_index)
{
f64 ipc_rate;
ASSERT (i + 1 < vec_len (c->counter_names));
ASSERT ((i + 1) < vec_len (c->counter_names));
if (c->counter_values[i + 1] > 0)
ipc_rate = (f64) c->counter_values[i]
src/plugins/perfmon/perfmon.h
+9 -5
View File
@@ -97,8 +97,11 @@ typedef struct
perfmon_cpuid_and_table_t *perfmon_tables;
uword *perfmon_table;
/* vector of events to collect */
perfmon_event_config_t *events_to_collect;
/* vector of single events to collect */
perfmon_event_config_t *single_events_to_collect;
/* vector of paired events to collect */
perfmon_event_config_t *paired_events_to_collect;
/* Base indices of synthetic event tuples */
u32 ipc_event_index;
@@ -109,13 +112,14 @@ typedef struct
/* Current event (index) being collected */
u32 current_event;
u32 *rdpmc_indices;
int n_active;
u32 **rdpmc_indices;
/* mmap base / size of (mapped) struct perf_event_mmap_page */
u8 **perf_event_pages;
u8 ***perf_event_pages;
u32 page_size;
/* Current perf_event file descriptors, per thread */
int *pm_fds;
int **pm_fds;
/* Logging */
vlib_log_class_t log_class;
src/plugins/perfmon/perfmon_periodic.c
+183 -122
View File
File diff suppressed because it is too large Load Diff
src/vlib/main.c
+48 -38
View File
@@ -543,15 +543,17 @@ never_inline void
vlib_node_runtime_sync_stats (vlib_main_t * vm,
vlib_node_runtime_t * r,
uword n_calls, uword n_vectors, uword n_clocks,
uword n_ticks)
uword n_ticks0, uword n_ticks1)
{
vlib_node_t *n = vlib_get_node (vm, r->node_index);
n->stats_total.calls += n_calls + r->calls_since_last_overflow;
n->stats_total.vectors += n_vectors + r->vectors_since_last_overflow;
n->stats_total.clocks += n_clocks + r->clocks_since_last_overflow;
n->stats_total.perf_counter_ticks += n_ticks +
r->perf_counter_ticks_since_last_overflow;
n->stats_total.perf_counter0_ticks += n_ticks0 +
r->perf_counter0_ticks_since_last_overflow;
n->stats_total.perf_counter1_ticks += n_ticks1 +
r->perf_counter1_ticks_since_last_overflow;
n->stats_total.perf_counter_vectors += n_vectors +
r->perf_counter_vectors_since_last_overflow;
n->stats_total.max_clock = r->max_clock;
@@ -560,7 +562,8 @@ vlib_node_runtime_sync_stats (vlib_main_t * vm,
r->calls_since_last_overflow = 0;
r->vectors_since_last_overflow = 0;
r->clocks_since_last_overflow = 0;
r->perf_counter_ticks_since_last_overflow = 0ULL;
r->perf_counter0_ticks_since_last_overflow = 0ULL;
r->perf_counter1_ticks_since_last_overflow = 0ULL;
r->perf_counter_vectors_since_last_overflow = 0ULL;
}
@@ -568,12 +571,12 @@ always_inline void __attribute__ ((unused))
vlib_process_sync_stats (vlib_main_t * vm,
vlib_process_t * p,
uword n_calls, uword n_vectors, uword n_clocks,
uword n_ticks)
uword n_ticks0, uword n_ticks1)
{
vlib_node_runtime_t *rt = &p->node_runtime;
vlib_node_t *n = vlib_get_node (vm, rt->node_index);
vlib_node_runtime_sync_stats (vm, rt, n_calls, n_vectors, n_clocks,
n_ticks);
n_ticks0, n_ticks1);
n->stats_total.suspends += p->n_suspends;
p->n_suspends = 0;
}
@@ -599,7 +602,7 @@ vlib_node_sync_stats (vlib_main_t * vm, vlib_node_t * n)
vec_elt_at_index (vm->node_main.nodes_by_type[n->type],
n->runtime_index);
vlib_node_runtime_sync_stats (vm, rt, 0, 0, 0, 0);
vlib_node_runtime_sync_stats (vm, rt, 0, 0, 0, 0, 0);
/* Sync up runtime next frame vector counters with main node structure. */
{
@@ -620,27 +623,30 @@ vlib_node_runtime_update_stats (vlib_main_t * vm,
vlib_node_runtime_t * node,
uword n_calls,
uword n_vectors, uword n_clocks,
uword n_ticks)
uword n_ticks0, uword n_ticks1)
{
u32 ca0, ca1, v0, v1, cl0, cl1, r;
u32 ptick0, ptick1, pvec0, pvec1;
u32 ptick00, ptick01, ptick10, ptick11, pvec0, pvec1;
cl0 = cl1 = node->clocks_since_last_overflow;
ca0 = ca1 = node->calls_since_last_overflow;
v0 = v1 = node->vectors_since_last_overflow;
ptick0 = ptick1 = node->perf_counter_ticks_since_last_overflow;
ptick00 = ptick01 = node->perf_counter0_ticks_since_last_overflow;
ptick10 = ptick11 = node->perf_counter1_ticks_since_last_overflow;
pvec0 = pvec1 = node->perf_counter_vectors_since_last_overflow;
ca1 = ca0 + n_calls;
v1 = v0 + n_vectors;
cl1 = cl0 + n_clocks;
ptick1 = ptick0 + n_ticks;
ptick01 = ptick00 + n_ticks0;
ptick11 = ptick10 + n_ticks1;
pvec1 = pvec0 + n_vectors;
node->calls_since_last_overflow = ca1;
node->clocks_since_last_overflow = cl1;
node->vectors_since_last_overflow = v1;
node->perf_counter_ticks_since_last_overflow = ptick1;
node->perf_counter0_ticks_since_last_overflow = ptick01;
node->perf_counter1_ticks_since_last_overflow = ptick11;
node->perf_counter_vectors_since_last_overflow = pvec1;
node->max_clock_n = node->max_clock > n_clocks ?
@@ -649,38 +655,39 @@ vlib_node_runtime_update_stats (vlib_main_t * vm,
r = vlib_node_runtime_update_main_loop_vector_stats (vm, node, n_vectors);
if (PREDICT_FALSE (ca1 < ca0 || v1 < v0 || cl1 < cl0) || (ptick1 < ptick0)
|| (pvec1 < pvec0))
if (PREDICT_FALSE (ca1 < ca0 || v1 < v0 || cl1 < cl0) || (ptick01 < ptick00)
|| (ptick11 < ptick10) || (pvec1 < pvec0))
{
node->calls_since_last_overflow = ca0;
node->clocks_since_last_overflow = cl0;
node->vectors_since_last_overflow = v0;
node->perf_counter_ticks_since_last_overflow = ptick0;
node->perf_counter0_ticks_since_last_overflow = ptick00;
node->perf_counter1_ticks_since_last_overflow = ptick10;
node->perf_counter_vectors_since_last_overflow = pvec0;
vlib_node_runtime_sync_stats (vm, node, n_calls, n_vectors, n_clocks,
n_ticks);
n_ticks0, n_ticks1);
}
return r;
}
static inline u64
vlib_node_runtime_perf_counter (vlib_main_t * vm)
static inline void
vlib_node_runtime_perf_counter (vlib_main_t * vm, u64 * pmc0, u64 * pmc1)
{
*pmc0 = 0;
*pmc1 = 0;
if (PREDICT_FALSE (vm->vlib_node_runtime_perf_counter_cb != 0))
return ((*vm->vlib_node_runtime_perf_counter_cb) (vm));
return 0ULL;
(*vm->vlib_node_runtime_perf_counter_cb) (vm, pmc0, pmc1);
}
always_inline void
vlib_process_update_stats (vlib_main_t * vm,
vlib_process_t * p,
uword n_calls, uword n_vectors, uword n_clocks,
uword n_ticks)
uword n_calls, uword n_vectors, uword n_clocks)
{
vlib_node_runtime_update_stats (vm, &p->node_runtime,
n_calls, n_vectors, n_clocks, n_ticks);
n_calls, n_vectors, n_clocks, 0ULL, 0ULL);
}
static clib_error_t *
@@ -1098,6 +1105,8 @@ dispatch_pcap_trace (vlib_main_t * vm,
}
}
u64 oingo0, oingo1;
static_always_inline u64
dispatch_node (vlib_main_t * vm,
vlib_node_runtime_t * node,
@@ -1146,18 +1155,14 @@ dispatch_node (vlib_main_t * vm,
if (1 /* || vm->thread_index == node->thread_index */ )
{
u64 pmc_before, pmc_delta;
u64 pmc_before[2], pmc_after[2], pmc_delta[2];
vlib_elog_main_loop_event (vm, node->node_index,
last_time_stamp,
frame ? frame->n_vectors : 0,
/* is_after */ 0);
/*
* To validate accounting: pmc_before = last_time_stamp
* perf ticks should equal clocks/pkt...
*/
pmc_before = vlib_node_runtime_perf_counter (vm);
vlib_node_runtime_perf_counter (vm, &pmc_before[0], &pmc_before[1]);
/*
* Turn this on if you run into
@@ -1191,7 +1196,10 @@ dispatch_node (vlib_main_t * vm,
* To validate accounting: pmc_delta = t - pmc_before;
* perf ticks should equal clocks/pkt...
*/
pmc_delta = vlib_node_runtime_perf_counter (vm) - pmc_before;
vlib_node_runtime_perf_counter (vm, &pmc_after[0], &pmc_after[1]);
pmc_delta[0] = pmc_after[0] - pmc_before[0];
pmc_delta[1] = pmc_after[1] - pmc_before[1];
vlib_elog_main_loop_event (vm, node->node_index, t, n, /* is_after */
1);
@@ -1199,11 +1207,18 @@ dispatch_node (vlib_main_t * vm,
vm->main_loop_vectors_processed += n;
vm->main_loop_nodes_processed += n > 0;
if (pmc_delta[0] || pmc_delta[1])
{
oingo0 += pmc_delta[0];
oingo1 += pmc_delta[1];
}
v = vlib_node_runtime_update_stats (vm, node,
/* n_calls */ 1,
/* n_vectors */ n,
/* n_clocks */ t - last_time_stamp,
pmc_delta /* PMC ticks */ );
pmc_delta[0] /* PMC0 */ ,
pmc_delta[1] /* PMC1 */ );
/* When in interrupt mode and vector rate crosses threshold switch to
polling mode. */
@@ -1542,8 +1557,7 @@ dispatch_process (vlib_main_t * vm,
vlib_process_update_stats (vm, p,
/* n_calls */ !is_suspend,
/* n_vectors */ n_vectors,
/* n_clocks */ t - last_time_stamp,
/* pmc_ticks */ 0ULL);
/* n_clocks */ t - last_time_stamp);
return t;
}
@@ -1626,8 +1640,7 @@ dispatch_suspended_process (vlib_main_t * vm,
vlib_process_update_stats (vm, p,
/* n_calls */ !is_suspend,
/* n_vectors */ n_vectors,
/* n_clocks */ t - last_time_stamp,
/* pmc_ticks */ 0ULL);
/* n_clocks */ t - last_time_stamp);
return t;
}
@@ -1677,9 +1690,6 @@ vlib_main_or_worker_loop (vlib_main_t * vm, int is_main)
if (!nm->interrupt_threshold_vector_length)
nm->interrupt_threshold_vector_length = 5;
/* Make sure the performance monitor counter is disabled */
vm->perf_counter_id = ~0;
/* Start all processes. */
if (is_main)
{
src/vlib/main.h
+2 -3
View File
@@ -84,9 +84,8 @@ typedef struct vlib_main_t
u32 node_counts_per_main_loop[2];
/* Main loop hw / sw performance counters */
u64 (*vlib_node_runtime_perf_counter_cb) (struct vlib_main_t *);
int perf_counter_id;
int perf_counter_fd;
void (*vlib_node_runtime_perf_counter_cb) (struct vlib_main_t *,
u64 *, u64 *);
/* Every so often we switch to the next counter. */
#define VLIB_LOG2_MAIN_LOOPS_PER_STATS_UPDATE 7
src/vlib/node.h
+4 -2
View File
@@ -258,7 +258,8 @@ typedef struct
u64 calls, vectors, clocks, suspends;
u64 max_clock;
u64 max_clock_n;
u64 perf_counter_ticks;
u64 perf_counter0_ticks;
u64 perf_counter1_ticks;
u64 perf_counter_vectors;
} vlib_node_stats_t;
@@ -507,7 +508,8 @@ typedef struct vlib_node_runtime_t
u32 vectors_since_last_overflow; /**< Number of vector elements
processed by this node. */
u32 perf_counter_ticks_since_last_overflow; /**< Perf counter ticks */
u32 perf_counter0_ticks_since_last_overflow; /**< Perf counter 0 ticks */
u32 perf_counter1_ticks_since_last_overflow; /**< Perf counter 1 ticks */
u32 perf_counter_vectors_since_last_overflow; /**< Perf counter vectors */
u32 next_frame_index; /**< Start of next frames for this
src/vlib/node_cli.c
-15
View File
@@ -148,8 +148,6 @@ format_vlib_node_stats (u8 * s, va_list * va)
f64 maxc, maxcn;
u32 maxn;
u32 indent;
u64 pmc_ticks;
f64 pmc_ticks_per_packet;
if (!n)
{
@@ -163,9 +161,6 @@ format_vlib_node_stats (u8 * s, va_list * va)
"%=30s%=12s%=16s%=16s%=16s%=16s%=16s",
"Name", "State", "Calls", "Vectors", "Suspends",
"Clocks", "Vectors/Call");
if (vm->perf_counter_id)
s = format (s, "%=16s", "Perf Ticks");
return s;
}
@@ -182,13 +177,6 @@ format_vlib_node_stats (u8 * s, va_list * va)
else
maxcn = 0.0;
pmc_ticks = n->stats_total.perf_counter_ticks -
n->stats_last_clear.perf_counter_ticks;
if (p > 0)
pmc_ticks_per_packet = (f64) pmc_ticks / (f64) p;
else
pmc_ticks_per_packet = 0.0;
/* Clocks per packet, per call or per suspend. */
x = 0;
if (p > 0)
@@ -221,9 +209,6 @@ format_vlib_node_stats (u8 * s, va_list * va)
s = format (s, "%-30v%=12U%16Ld%16Ld%16Ld%16.2e%16.2f", ns,
format_vlib_node_state, vm, n, c, p, d, x, v);
if (pmc_ticks_per_packet > 0.0)
s = format (s, "%16.2e", pmc_ticks_per_packet);
if (ns != n->name)
vec_free (ns);
src/vlibapi/node_serialize.c
+2 -9
View File
@@ -57,7 +57,7 @@ vlib_node_serialize (vlib_main_t * vm, vlib_node_t *** node_dups, u8 * vector,
u8 *namep;
u32 name_bytes;
uword i, j, k;
u64 l, v, c, d, pmc;
u64 l, v, c, d;
state_string_enum_t state_code;
serialize_open_vector (sm, vector);
@@ -77,8 +77,6 @@ vlib_node_serialize (vlib_main_t * vm, vlib_node_t *** node_dups, u8 * vector,
v = n->stats_total.vectors - n->stats_last_clear.vectors;
c = n->stats_total.calls - n->stats_last_clear.calls;
d = n->stats_total.suspends - n->stats_last_clear.suspends;
pmc = n->stats_total.perf_counter_ticks
- n->stats_last_clear.perf_counter_ticks;
state_code = STATE_INTERNAL;
@@ -151,8 +149,6 @@ vlib_node_serialize (vlib_main_t * vm, vlib_node_t *** node_dups, u8 * vector,
serialize_integer (sm, v, 8);
/* Total suspends */
serialize_integer (sm, d, 8);
/* PMC counter */
serialize_integer (sm, pmc, 8);
}
else /* no stats */
serialize_likely_small_unsigned_integer (sm, 0);
@@ -171,7 +167,7 @@ vlib_node_unserialize (u8 * vector)
vlib_node_t **nodes;
vlib_node_t ***nodes_by_thread = 0;
int i, j, k;
u64 l, v, c, d, pmc;
u64 l, v, c, d;
state_string_enum_t state_code;
int stats_present;
@@ -229,9 +225,6 @@ vlib_node_unserialize (u8 * vector)
/* Total suspends */
unserialize_integer (sm, &d, 8);
node->stats_total.suspends = d;
/* PMC counter */
unserialize_integer (sm, &pmc, 8);
node->stats_total.perf_counter_ticks = pmc;
}
}
}