[
https://issues.apache.org/jira/browse/SYSTEMML-1047?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Mike Dusenberry updated SYSTEMML-1047:
--------------------------------------
Description:
Currently, when MLContext is run with Statistics turned on, the resulting
statistics are inaccurate. In particular, the 3 measures of elapsed time are
set to "0.000 seconds" which is obviously inaccurate. Furthermore, after
{{MLContext}} is used for the first script with stats on, subsequent scripts
run with stats turned on will no longer provide updated statistics, and will
simply print out the original stats.
Example:
{code}
SystemML Statistics:
Total elapsed time: 0.000 sec.
Total compilation time: 0.000 sec.
Total execution time: 0.000 sec.
Number of compiled Spark inst: 164.
Number of executed Spark inst: 118.
Cache hits (Mem, WB, FS, HDFS): 1626/0/0/0.
Cache writes (WB, FS, HDFS): 679/0/0.
Cache times (ACQr/m, RLS, EXP): 0.002/0.001/3.998/0.000 sec.
HOP DAGs recompiled (PRED, SB): 0/43.
HOP DAGs recompile time: 0.051 sec.
Functions recompiled: 6.
Functions recompile time: 0.037 sec.
Spark ctx create time (lazy): 0.000 sec.
Spark trans counts (par,bc,col):0/0/1.
Spark trans times (par,bc,col): 0.000/0.000/0.577 secs.
Total JIT compile time: 135.184 sec.
Total JVM GC count: 1326.
Total JVM GC time: 160.557 sec.
Heavy hitter instructions (name, time, count):
-- 1) train 59.288 sec 1
-- 2) predict 23.537 sec 5
-- 3) conv2d 14.528 sec 42
-- 4) sel+ 11.446 sec 56
-- 5) * 9.014 sec 330
-- 6) + 5.137 sec 343
-- 7) > 3.841 sec 16
-- 8) conv2d_backward_filter 3.008 sec 12
-- 9) conv2d_backward_data 2.378 sec 8
-- 10) maxpooling_backward 2.056 sec 12
{code}
cc [~deron]
was:
Currently, when MLContext is run with Statistics turned on, the resulting
statistics are inaccurate. In particular, the 3 measures of elapsed time are
set to "0.000 seconds" which is obviously inaccurate.
Example:
{code}
SystemML Statistics:
Total elapsed time: 0.000 sec.
Total compilation time: 0.000 sec.
Total execution time: 0.000 sec.
Number of compiled Spark inst: 164.
Number of executed Spark inst: 118.
Cache hits (Mem, WB, FS, HDFS): 1626/0/0/0.
Cache writes (WB, FS, HDFS): 679/0/0.
Cache times (ACQr/m, RLS, EXP): 0.002/0.001/3.998/0.000 sec.
HOP DAGs recompiled (PRED, SB): 0/43.
HOP DAGs recompile time: 0.051 sec.
Functions recompiled: 6.
Functions recompile time: 0.037 sec.
Spark ctx create time (lazy): 0.000 sec.
Spark trans counts (par,bc,col):0/0/1.
Spark trans times (par,bc,col): 0.000/0.000/0.577 secs.
Total JIT compile time: 135.184 sec.
Total JVM GC count: 1326.
Total JVM GC time: 160.557 sec.
Heavy hitter instructions (name, time, count):
-- 1) train 59.288 sec 1
-- 2) predict 23.537 sec 5
-- 3) conv2d 14.528 sec 42
-- 4) sel+ 11.446 sec 56
-- 5) * 9.014 sec 330
-- 6) + 5.137 sec 343
-- 7) > 3.841 sec 16
-- 8) conv2d_backward_filter 3.008 sec 12
-- 9) conv2d_backward_data 2.378 sec 8
-- 10) maxpooling_backward 2.056 sec 12
{code}
cc [~deron]
> Statistics w/ MLContext Yields Incorrect Results
> ------------------------------------------------
>
> Key: SYSTEMML-1047
> URL: https://issues.apache.org/jira/browse/SYSTEMML-1047
> Project: SystemML
> Issue Type: Bug
> Reporter: Mike Dusenberry
>
> Currently, when MLContext is run with Statistics turned on, the resulting
> statistics are inaccurate. In particular, the 3 measures of elapsed time are
> set to "0.000 seconds" which is obviously inaccurate. Furthermore, after
> {{MLContext}} is used for the first script with stats on, subsequent scripts
> run with stats turned on will no longer provide updated statistics, and will
> simply print out the original stats.
> Example:
> {code}
> SystemML Statistics:
> Total elapsed time: 0.000 sec.
> Total compilation time: 0.000 sec.
> Total execution time: 0.000 sec.
> Number of compiled Spark inst: 164.
> Number of executed Spark inst: 118.
> Cache hits (Mem, WB, FS, HDFS): 1626/0/0/0.
> Cache writes (WB, FS, HDFS): 679/0/0.
> Cache times (ACQr/m, RLS, EXP): 0.002/0.001/3.998/0.000 sec.
> HOP DAGs recompiled (PRED, SB): 0/43.
> HOP DAGs recompile time: 0.051 sec.
> Functions recompiled: 6.
> Functions recompile time: 0.037 sec.
> Spark ctx create time (lazy): 0.000 sec.
> Spark trans counts (par,bc,col):0/0/1.
> Spark trans times (par,bc,col): 0.000/0.000/0.577 secs.
> Total JIT compile time: 135.184 sec.
> Total JVM GC count: 1326.
> Total JVM GC time: 160.557 sec.
> Heavy hitter instructions (name, time, count):
> -- 1) train 59.288 sec 1
> -- 2) predict 23.537 sec 5
> -- 3) conv2d 14.528 sec 42
> -- 4) sel+ 11.446 sec 56
> -- 5) * 9.014 sec 330
> -- 6) + 5.137 sec 343
> -- 7) > 3.841 sec 16
> -- 8) conv2d_backward_filter 3.008 sec 12
> -- 9) conv2d_backward_data 2.378 sec 8
> -- 10) maxpooling_backward 2.056 sec 12
> {code}
> cc [~deron]
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
