stats.c File Reference

Statistics module. More...

#include <log/stats.h>
#include <arch/io.h>
#include <arch/mem.h>
#include <arch/timer.h>
#include <distributed/mpi.h>
#include <log/file.h>
#include <mm/mm.h>
#include <assert.h>
#include <stdint.h>
#include <stdio.h>

Include dependency graph for stats.c:

Classes
struct	stats_thread
	A container for statistics in a logical time period. More...
struct	stats_node
	A container for node-wide statistics in a logical time period. More...
struct	stats_global
	A container for node-wide statistics relevant for the whole simulation run. More...

Macros
#define	STATS_BUFFER_ENTRIES   (1024)
	The number of entries to keep in the stdio buffers before flushing a temporary statistics file to disk.

Functions
void	stats_global_time_take (enum stats_global_type this_stat)
	Take a lifetime event time value. More...
void	stats_global_init (void)
	Initializes the stats subsystem in the node.
void	stats_init (void)
	Initializes the stats subsystem in the current thread.
static void	stats_files_receive (FILE *out_f)
	Collect and dump the statistics from other nodes in the final binary file. More...
static void	stats_files_send (void)
	Send the final statistics data of this node to the master node.
static void	stats_file_final_write (FILE *out_f)
	Write the final statistic file, except for the data from other nodes. More...
void	stats_global_fini (void)
	Finalizes the stats subsystem in the node. More...
void	stats_take (enum stats_thread_type this_stat, uint_fast64_t c)
	Sum a sample to a statistics value. More...
void	stats_on_gvt (simtime_t gvt)
	Perform GVT related activities for the statistics subsystem. More...
void	stats_dump (void)
	Dump some final minimal statistics on screen.
uint64_t	stats_retrieve (enum stats_thread_type this_stat)
	Retrieve the value of a metric of this thread. More...

Variables
const char *const	stats_names []
	The statistics names, used to fill in the preamble of the final statistics binary file. More...
static timer_uint	sim_start_ts
	The first timestamp ever collected for this simulation run.
static timer_uint	sim_start_ts_hr
	The first high resolution timestamp of this simulation run: used to correlate high resolution and wall clock timers.
static struct stats_global	stats_glob_cur
	The global stats.
static FILE *	stats_node_tmp
	A pointer to the temporary file used to save stats_node structs produced during simulation.
static FILE **	stats_tmps
	An array of pointers to the temporary files used to save stats_thread structs produced by threads during simulation.
static __thread struct stats_thread	stats_cur
	The current values of thread statistics for this logical time period (from the previous GVT to the next one)

Detailed Description

Statistics module.

All facilities to collect, gather, and dump statistics are implemented in this module.

Copyright

Copyright (C) 2008-2022 HPDCS Group https://hpdcs.github.io

Function Documentation

stats_file_final_write()

static void stats_file_final_write ( FILE *  out_f )

static

Write the final statistic file, except for the data from other nodes.

Parameters

[out]

out_f

a pointer to the file to write

The following tables document the content of the final generated statistics file.

• Rows are in order of appearance in the binary file
• The binary file starts with a single Preamble
• The Count column specifies how many repeated elements of that row are present
• The Size column reports the size of a single row element in bytes
• The Type column specifies the type of the field: uppercase types are further described in their own table
• The Ref column assigns an identifier for values that are used later on to define counts of other rows
• The Description column documents the purpose of the field

Preamble:

Count	Size	Type	Ref	Description
1	2	uint	–	Magic number used to detect the endianness of the machine
1	8	int	s_cnt	Count of available thread metrics
s_cnt	*	Pascal string	–	Names of the thread metrics
1	8	int	n_cnt	Count of MPI ranks (1 in the case of a single node run)
n_cnt	*	Node stats	–	The statistics produced by the nodes

Pascal string:

Count	Size	Type	Ref	Description
1	1	uint	p_len	Size of the Pascal string (maximum 255, clearly)
p_len	1	char	–	Content of the string

Node stats:

Count	Size	Type	Ref	Description
1	8	uint	t_cnt	Count of threads for the this node
1	8	uint	–	Maximum resident set size of this node (in bytes)
6	8	uint	–	Some timestamps in us (see enum stats_global_type)
1	8	uint	–	High resolution time end to end value (see enum stats_global_type)
1	8	int	n_siz	Size of the node GVT stats array
n_siz / 16	16	Node GVT entry	–	The node-wide statistics produced at each GVT by this node
t_cnt	*	Thread GVT stats	–	The statistics produced by each thread on this node

Node GVT entry:

Count	Size	Type	Ref	Description
1	8	double	–	Global virtual time value
1	8	uint	–	Current resident set size of this node (in bytes)

Thread stats:

Count	Size	Type	Ref	Description
1	8	uint	t_siz	Size in bytes of this thread's GVT stats array
t_siz / (s_cnt * 8)	s_cnt * 8	Thread GVT entry	–	The statistics produced at each GVT by this thread

Thread GVT entry:

Count	Size	Typ	Ref	Description
s_cnt	8	uint	–	The values of the stats described in the Preamble for this GVT and thread

In a correctly completed simulation n_siz / 16 == t_siz / (s_cnt * 8) for each node and thread (this is the number of committed GVTs). This function only writes the Node stats for the current node (the master node in a MPI run). The function stats_files_receive() deals with the other nodes. TODO add to the file other kind of statistics, for example ROOT-Sim config, machine hardware etc

stats_files_receive()

static void stats_files_receive ( FILE *  out_f )

static

Collect and dump the statistics from other nodes in the final binary file.

Parameters

[out]

out_f

a pointer to the file to write. This can be NULL: the data from other nodes won't be saved

You may want to call this function with out_f set to NULL to flush the pending MPI communications. For details about the binary file format see stats_file_final_write().

stats_global_fini()

void stats_global_fini

(

void

)

Finalizes the stats subsystem in the node.

When finalizing this subsystem, the master node dumps his statistics from his temporary files onto the final binary file. Then, in a distributed setting, he receives the slaves temporary files, dumping their statistics as well.

stats_global_time_take()

void stats_global_time_take

(

enum stats_global_type

this_stat

)

Take a lifetime event time value.

Parameters

this_stat

The type of event just occurred

stats_on_gvt()

void stats_on_gvt

(

simtime_t

gvt

)

Perform GVT related activities for the statistics subsystem.

Parameters

gvt	the time value of the current GVT

Dumps accumulated statistics to the file and resets the statistics buffer to ready up for the following processing phase

stats_retrieve()

uint64_t stats_retrieve

(

enum stats_thread_type

this_stat

)

Retrieve the value of a metric of this thread.

This values are computed since the end of the last GVT.

stats_take()

void stats_take	(	enum stats_thread_type	this_stat,
		uint_fast64_t	c
	)

Sum a sample to a statistics value.

Parameters

this_stat	the statistics type to add the sample to
c	the sample to sum

Variable Documentation

stats_names

const char* const stats_names[]

Initial value:

= {
    [STATS_MSG_PROCESSED] = "processed messages",
    [STATS_MSG_PROCESSED_TIME] = "processed messages time",
    [STATS_ROLLBACK] = "rollbacks",
    [STATS_RECOVERY_TIME] = "recovery time",
    [STATS_MSG_ROLLBACK] = "rolled back messages",
    [STATS_CKPT] = "checkpoints",
    [STATS_CKPT_TIME] = "checkpoints time",
    [STATS_CKPT_SIZE] = "checkpoints size",
    [STATS_MSG_SILENT] = "silent messages",
    [STATS_MSG_SILENT_TIME] = "silent messages time",
    [STATS_MSG_ANTI] = "anti messages",
    [STATS_REAL_TIME_GVT] = "gvt real time"
}

The statistics names, used to fill in the preamble of the final statistics binary file.