{"id":1061,"date":"2017-10-11T08:02:57","date_gmt":"2017-10-11T08:02:57","guid":{"rendered":"http:\/\/blog.tiran.info\/?p=1061"},"modified":"2017-10-11T08:02:57","modified_gmt":"2017-10-11T08:02:57","slug":"annore-4","status":"publish","type":"post","link":"https:\/\/blog.tiran.stream\/?p=1061","title":{"rendered":"ANN\/ORE #4 – Performance de neuralnet"},"content":{"rendered":"

On a vu dans le billet pr\u00e9c\u00e9dent<\/a> que la construction du mod\u00e8le avec le package neuralnet<\/a> \u00e9tait consid\u00e9rablement plus rapide qu’avec nnet<\/a>.<\/p>\n

Cette diff\u00e9rence m’a intrigu\u00e9 dans la mesure ou les deux packages doivent fonctionner – \u00e0 priori – selon une logique similaire: gradient de descente, forward et back propagation…<\/p>\n

En observant le profil de charge de la machine \u00e0 l’aide de la commande top pendant la phase de construction du mod\u00e8le, j’ai rapidement compris l’origine de la diff\u00e9rence. En effet, il appara\u00eet que l’impl\u00e9mentation de neuralnet est multithread\u00e9<\/a> contrairement \u00e0 celle de nnet:<\/p>\n

\"\"<\/a><\/p>\n

 <\/p>\n

 <\/p>\n

Dans un contexte ORE, cela signifie que le processus extproc instancie plusieurs threads qui op\u00e8rent en parall\u00e8le. Le champ int\u00e9ressant ici est nlwp<\/a> qui indique le nombre de threads du processus extproc:<\/p>\n

\"\"<\/a><\/p>\n

 <\/p>\n

 <\/p>\n

 <\/p>\n

Le script de collecte (rudimentaire!) suivant a donc \u00e9t\u00e9 lanc\u00e9 pendant la construction des mod\u00e8les avec nnet et neuralnet (fichiers extproc_nnet<\/a> et extproc_neuralnet<\/a>).<\/p>\n

$ cat extproc_ps.sh\n#!\/bin\/bash\n\n>extproc.txt\n\nwhile true\ndo\n        date >>extproc.txt\n        ps -C extproc -o pid,rss,vsz,time,nlwp >>extproc.txt\n        sleep 30\ndone\n\n$\n<\/pre>\n
 <\/p>\n
Les donn\u00e9es collect\u00e9es ont \u00e9t\u00e9 exploit\u00e9es \u00e0 l’aide d’une table externe:<\/p>\n
SQL> CREATE OR REPLACE DIRECTORY d1 AS '\/tmp';\n\nDirectory created.\n\nSQL>\nSQL> CREATE TABLE exttab_extproc\n  2  (\n  3      line VARCHAR2 (300)\n  4  )\n  5  ORGANIZATION EXTERNAL\n  6      (TYPE oracle_loader\n  7       DEFAULT DIRECTORY D1\n  8       ACCESS PARAMETERS (\n  9           RECORDS DELIMITED BY NEWLINE\n 10       )\n 11       LOCATION ('extproc_neuralnet.txt'));\n\nTable created.\n\nSQL>\nSQL>\nSQL> SELECT * FROM exttab_extproc FETCH FIRST 5 ROWS ONLY;\n\nLINE\n----------------------------------------------------------\nTue Sep 12 11:21:35 CEST 2017\n  PID   RSS    VSZ     TIME NLWP\n 9161 51024 362776 00:00:00    1\nTue Sep 12 11:22:05 CEST 2017\n  PID   RSS    VSZ     TIME NLWP\n\nSQL>\n<\/pre>\n
La mise en forme est ensuite r\u00e9alis\u00e9e \u00e0 l’aide de d’expressions r\u00e9guli\u00e8res<\/a> et de CTE<\/a>. Le r\u00e9sultat est pr\u00e9sent\u00e9 via une vue:<\/p>\n
SQL> CREATE OR REPLACE VIEW evol_process_extproc\n  2  AS\n  3      WITH\n  4          procres\n  5          AS\n  6              (SELECT ROWNUM                                        rn,\n  7                      REGEXP_REPLACE (line, '[[:space:]]{2,}', ' ') line\n  8                 FROM exttab_extproc\n  9                WHERE line NOT LIKE '%PID%RSS%'),\n 10          procres_evol\n 11          AS\n 12              (SELECT line,\n 13                      TRIM (LEAD (line) OVER (PARTITION BY NULL ORDER BY rn))\n 14                          next_line\n 15                 FROM procres)\n 16      SELECT TO_DATE (REPLACE (line, ' CEST ', ' '),\n 17                      'Dy Mon DD HH24:MI:SS YYYY')\n 18                 dt,\n 19             REGEXP_SUBSTR (next_line,\n 20                            '(.*?)([[:space:]]|$)',\n 21                            1,\n 22                            1,\n 23                            NULL,\n 24                            1)\n 25                 pid,\n 26             REGEXP_SUBSTR (next_line,\n 27                            '(.*?)([[:space:]]|$)',\n 28                            1,\n 29                            2,\n 30                            NULL,\n 31                            1)\n 32                 rss,\n 33             REGEXP_SUBSTR (next_line,\n 34                            '(.*?)([[:space:]]|$)',\n 35                            1,\n 36                            3,\n 37                            NULL,\n 38                            1)\n 39                 vsz,\n 40               (  TO_DATE (REGEXP_SUBSTR (next_line,\n 41                                          '(.*?)([[:space:]]|$)',\n 42                                          1,\n 43                                          4,\n 44                                          NULL,\n 45                                          1),\n 46                           'HH24:MI:SS')\n 47                - TRUNC (TO_DATE ('00:00', 'HH24:MI')))\n 48             * 86400\n 49                 cpusec,\n 50             REGEXP_SUBSTR (next_line,\n 51                            '(.*?)([[:space:]]|$)',\n 52                            1,\n 53                            5,\n 54                            NULL,\n 55                            1)\n 56                 nlwp\n 57        FROM procres_evol\n 58       WHERE next_line NOT LIKE '%2017';\n\nView created.\n\nSQL>\nSQL> column NLWP format a5\nSQL> column PID format a8\nSQL> column RSS format a10\nSQL> column VSZ format a10\nSQL>\nSQL> SELECT * FROM evol_process_extproc FETCH FIRST 10 ROWS ONLY;\n\nDT                PID      RSS        VSZ            CPUSEC NLWP\n----------------- -------- ---------- ---------- ---------- -----\n12\/09\/17 11:21:35 9161     51024      362776              0 1\n12\/09\/17 11:22:05 9161     51024      362776              0 1\n12\/09\/17 11:22:35 9161     51024      362776              0 1\n12\/09\/17 11:23:05 9161     51024      362776              0 1\n12\/09\/17 11:23:35 9161     51024      362776              0 1\n12\/09\/17 11:24:05 9161     200128     529856              7 1\n12\/09\/17 11:24:35 9161     1803592    2132512            35 1\n12\/09\/17 11:25:05 9161     2869720    4926496           164 25\n12\/09\/17 11:25:35 9161     3903716    5960336           499 25\n12\/09\/17 11:26:05 9161     3676520    5733096           785 25\n\n10 rows selected.\n\nSQL>\n<\/pre>\n
Finalement, la requ\u00eate suivante (bas\u00e9e sur des fonctions analytiques<\/a>) est utilis\u00e9e pour calculer le nombre de threads actifs (champ AAT) par p\u00e9riode d’\u00e9chantillonnage:<\/p>\n
SQL>   SELECT dt,\n  2           nlwp,\n  3           vsz,\n  4           rss,\n  5           ROUND (cpusec \/ elaps, 1) aat\n  6      FROM (SELECT dt,\n  7                   86400 * (dt - LAG (dt) OVER (PARTITION BY NULL ORDER BY dt))\n  8                       elaps,\n  9                   cpusec - LAG (cpusec) OVER (PARTITION BY NULL ORDER BY dt)\n 10                       cpusec,\n 11                   nlwp,\n 12                   vsz,\n 13                   rss\n 14              FROM evol_process_extproc)\n 15  ORDER BY dt\n 16  FETCH FIRST 10 ROWS ONLY;\n\nDT                NLWP  VSZ        RSS               AAT\n----------------- ----- ---------- ---------- ----------\n12\/09\/17 11:21:35 1     362776     51024\n12\/09\/17 11:22:05 1     362776     51024               0\n12\/09\/17 11:22:35 1     362776     51024               0\n12\/09\/17 11:23:05 1     362776     51024               0\n12\/09\/17 11:23:35 1     362776     51024               0\n12\/09\/17 11:24:05 1     529856     200128             .2\n12\/09\/17 11:24:35 1     2132512    1803592            .9\n12\/09\/17 11:25:05 25    4926496    2869720           4.3\n12\/09\/17 11:25:35 25    5960336    3903716          11.2\n12\/09\/17 11:26:05 25    5733096    3676520           9.5\n\n10 rows selected.\n\nSQL>\n<\/pre>\n
On peut alors repr\u00e9senter \u00e0 l’aide de ggplot2<\/a>, le profil d’activit\u00e9 des threads lors de la construction du mod\u00e8le:<\/p>\n
> library(ROracle)\nLoading required package: DBI\n> ora = Oracle()\n> cnx = dbConnect(ora, username="c##rafa", password="Password1#", dbname="\/\/clorai2-scan:1521\/pdb_hodba08")\n> evol_cpu_mem <- dbGetQuery(cnx, "  SELECT dt,\n+          nlwp,\n+          vsz,\n+          rss,\n+          ROUND (cpusec \/ elaps, 1) aat\n+     FROM (SELECT dt,\n+                  86400 * (dt - LAG (dt) OVER (PARTITION BY NULL ORDER BY dt))\n+                      elaps,\n+                  cpusec - LAG (cpusec) OVER (PARTITION BY NULL ORDER BY dt)\n+                      cpusec,\n+                  nlwp,\n+                  vsz,\n+                  rss\n+             FROM evol_process_extproc)\n+ ORDER BY dt")\n> \n> library(ggplot2)\n> \n> evol_cpu_mem$couleur <- ifelse(evol_cpu_mem$AAT==0, "blue", \n+                                ifelse(evol_cpu_mem$AAT<0.7, "forestgreen",\n+                                       ifelse(evol_cpu_mem$AAT == 1, "darkorange", "red")\n+                                       )\n+                                )\n> \n> ggplot(evol_cpu_mem[-1,], aes(DT, AAT)) +\n+   geom_point(color=evol_cpu_mem$couleur[-1]) +\n+   geom_line(color=evol_cpu_mem$couleur[-1])  +   \n+   xlab("Heure") + \n+   ylab("Nombre moyen de threads actifs") +   \n+   theme_classic() +   \n+   ggtitle("Activit\u00e9 des threads lors de la construction du mod\u00e8le") +\n+   theme(plot.title = element_text(hjust = 0.5)) +\n+   theme(panel.grid.major = element_line(linetype = "dotted")) + \n+   scale_y_continuous(breaks = seq(0, 12),limits = c(0,11.5))\n> \n<\/pre>\n
Le graphique obtenu est coloris\u00e9 en fonction du niveau d’activit\u00e9 des threads. On constate que le profil n’est pas uniforme, en particulier ce n’est pas la totalit\u00e9 de la construction du mod\u00e8le qui donne lieu \u00e0 une activit\u00e9 multithread\u00e9e :<\/p>\n