[ACCUMULO-2749] 1.6.0RC4 slower than 1.5.1

• Type: Bug
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Fix Version/s: 1.6.0
• Assignee: Eric C. Newton
• Reporter: Keith Turner
• Created: 28/Apr/14 22:47
• Updated: 01/May/14 05:25
• Resolved: 29/Apr/14 16:24

Description

I am seeing consistently slower write rates when comparing 1.5.1 and 1.6.0 RC4. Following env :

• hadoop 2.2.0
• Zookeeper 3.4.5
• Centos 6 (native, not a VM)
• using examples/3GB/native-standalone config
• setting tserver.mutation.queue.max 4M
• I am running test/system/test1/ingest_test.sh
• single node

Saw the following rates

  1.5.1 aggregate rate : 200,806 records/sec
  1.6.0 aggregate rate : 181,264 records/sec (90%)

Reran test setting table.walog.enabled=false

  1.5.1 aggregate rate : 302,307 records/sec
  1.6.0 aggregate rate : 232,685 records/sec (77%)

Since the relative performance gap seemed to widen when disabling walog I tried another test where I disabled walogs and set the devnull iterator

  1.5.1 aggregate rate : 554,853 records/sec
  1.6.0 aggregate rate : 373,556 records/sec (67%)

The relative performance gap grew even further.

Used the following command to enable dev null iterator in 1.5.1 and RC4

config -t test_ingest -s table.iterator.minc.devnull=10,org.apache.accumulo.core.iterators.DevNull

Issue Links

• relates to: ACCUMULO-2762 Use better compiler optimizations by default for native maps