[
https://issues.apache.org/jira/browse/HADOOP-15554?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Todd Lipcon updated HADOOP-15554:
---------------------------------
Status: Patch Available (was: Open)
Attached patch refactors out the config-parsing code to a new inner class with
a bunch of smaller functions which are easier to compile. I also took the
opportunity to make a few micro-optimizations like avoiding construction of the
confSources array in the common case that the config file uses no "<source>"
tags.
I tested the improvement by running:
{code}
for x in $(seq 1 60); do
java -XX:+CITime -cp
hadoop-common-project/hadoop-common/target/hadoop-common-3.2.0-SNAPSHOT.jar:$CP
\
org.apache.hadoop.examples.ExampleDriver pi 1 1 2>&1 \
| grep 'Total comp'
done | tee /tmp/patch.txt
{code}
to measure the total compilation time in a simple LocalJobRunner MR job. I
grepped out the times and ran a t-test using R:
{code}
data: d.orig and d.patched
t = 36.511, df = 110.1, p-value < 2.2e-16
alternative hypothesis: true difference in means is not equal to 0
95 percent confidence interval:
0.7329980 0.8171354
sample estimates:
mean of x mean of y
3.508300 2.733233
{code}
So this saves about 730-810ms of CPU time spent by the JIT.
To test throughput, I used the ConfTest.java program from HADOOP-14216.
{code}
orig:
duration: 20745 count: 3561000
real 0m21.104s
user 0m29.296s
sys 0m1.903s
patch:
duration: 21810 count: 3561000
real 0m22.304s
user 0m27.013s
sys 0m2.547s
{code}
So it seems around the same - a bit less user time, a bit longer real time.
Close enough to call "not a regression".
I also tried 'fs -ls hdfs://nn/' under 'perf stat -r10':
{code}
orig:
5295.930635 task-clock (msec) # 3.454 CPUs utilized
( +- 3.56% )
10,977 context-switches # 0.002 M/sec
( +- 0.37% )
613 cpu-migrations # 0.116 K/sec
( +- 2.28% )
86,804 page-faults # 0.016 M/sec
( +- 0.12% )
14,823,251,627 cycles # 2.799 GHz
( +- 3.61% )
11,367,265,626 instructions # 0.77 insn per cycle
( +- 1.81% )
2,503,093,507 branches # 472.645 M/sec
( +- 3.26% )
67,066,880 branch-misses # 2.68% of all branches
( +- 0.23% )
1.533354188 seconds time elapsed
( +- 0.54% )
patch:
5173.366209 task-clock (msec) # 3.384 CPUs utilized
( +- 3.60% )
11,160 context-switches # 0.002 M/sec
( +- 1.32% )
630 cpu-migrations # 0.122 K/sec
( +- 2.82% )
87,732 page-faults # 0.017 M/sec
( +- 0.18% )
14,495,009,185 cycles # 2.802 GHz
( +- 3.55% )
11,485,553,655 instructions # 0.79 insn per cycle
( +- 1.80% )
2,487,385,519 branches # 480.806 M/sec
( +- 3.34% )
68,583,976 branch-misses # 2.76% of all branches
( +- 0.25% )
1.528788291 seconds time elapsed
( +- 0.62% )
{code}
and 'yarn application -list' on an RM running no applications:
{code}
orig:
2150.752819 task-clock (msec) # 2.101 CPUs utilized
( +- 0.89% )
9,179 context-switches # 0.004 M/sec
( +- 0.66% )
476 cpu-migrations # 0.221 K/sec
( +- 3.20% )
46,036 page-faults # 0.021 M/sec
( +- 0.13% )
5,928,445,661 cycles # 2.756 GHz
( +- 0.98% )
6,382,601,882 instructions # 1.08 insn per cycle
( +- 0.61% )
1,153,880,261 branches # 536.501 M/sec
( +- 0.60% )
47,370,186 branch-misses # 4.11% of all branches
( +- 0.65% )
1.023657616 seconds time elapsed
( +- 0.59% )
patch:
2106.716373 task-clock (msec) # 2.091 CPUs utilized
( +- 0.70% )
9,113 context-switches # 0.004 M/sec
( +- 0.62% )
451 cpu-migrations # 0.214 K/sec
( +- 1.46% )
47,218 page-faults # 0.022 M/sec
( +- 0.09% )
5,769,853,936 cycles # 2.739 GHz
( +- 0.73% )
6,320,641,188 instructions # 1.10 insn per cycle
( +- 0.31% )
1,141,174,880 branches # 541.684 M/sec
( +- 0.31% )
46,945,771 branch-misses # 4.11% of all branches
( +- 0.40% )
1.007474613 seconds time elapsed
( +- 0.50% )
{code}
So it seems a slight saving in cycles for both of those applications.
> Improve JIT performance for Configuration parsing
> -------------------------------------------------
>
> Key: HADOOP-15554
> URL: https://issues.apache.org/jira/browse/HADOOP-15554
> Project: Hadoop Common
> Issue Type: Improvement
> Components: conf, performance
> Affects Versions: 3.0.0
> Reporter: Todd Lipcon
> Assignee: Todd Lipcon
> Priority: Minor
> Attachments: hadoop-15554.patch
>
>
> In investigating a performance regression for small tasks between Hadoop 2
> and Hadoop 3, we found that the amount of time spent in JIT was significantly
> higher. Using jitwatch we were able to determine that, due to a combination
> of switching from DOM to SAX style parsing and just having more configuration
> key/value pairs, Configuration.loadResource is now getting compiled with the
> C2 compiler and taking quite some time. Breaking that very large function up
> into several smaller ones and eliminating some redundant bits of code
> improves the JIT performance measurably.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
