aurora-reviews mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Mehrdad Nurolahzade <mehr...@nurolahzade.com>
Subject Re: Review Request 55357: AURORA-1867 Consider reserving for multiple tasks per preemption round
Date Tue, 24 Jan 2017 16:42:00 GMT

-----------------------------------------------------------
This is an automatically generated e-mail. To reply, visit:
https://reviews.apache.org/r/55357/
-----------------------------------------------------------

(Updated Jan. 24, 2017, 8:42 a.m.)


Review request for Aurora, David McLaughlin, Joshua Cohen, Stephan Erb, and Zameer Manji.


Changes
-------

- Corrected spelling mistake
- Fixed bug in unit test


Bugs: AURORA-1867
    https://issues.apache.org/jira/browse/AURORA-1867


Repository: aurora


Description
-------

To be fair, PendingTaskProcessor interleaves tasks from different groups. However, this fairness
comes at the price of increasing reservation time. Even if reservations are being made for
the same task group, the processor would still restart iterating through slaves for each task
instance. This results in reevaluating all slaves already rejected in a previous search before
it finds a new viable candidate.

This patch improves `PendingTaskProcessor` performance by reducing slave search/evaluation
time, at the cost of reduced fairness. `PendingTaskProcessor` now does reservation for a configurable
maximum of _N_ candidates per task group in each iteration over the list of slaves.


Diffs (updated)
-----

  src/jmh/java/org/apache/aurora/benchmark/SchedulingBenchmarks.java fa37236e68657b539b182519b9d46d96d5b0953a

  src/main/java/org/apache/aurora/scheduler/preemptor/PendingTaskProcessor.java f59f3fd8959b1ba3726b55a2943fb9228a049ac5

  src/main/java/org/apache/aurora/scheduler/preemptor/PreemptorMetrics.java 67822cafbe89f4798b4ea6da3856663cc4872798

  src/main/java/org/apache/aurora/scheduler/preemptor/PreemptorModule.java 23d1c120657d5cb9d294a80c63e8a04512d361ca

  src/test/java/org/apache/aurora/scheduler/preemptor/PendingTaskProcessorTest.java d11ae5883f2a00dca4c4b36f0ab58ea95c7ecb2e

  src/test/java/org/apache/aurora/scheduler/preemptor/PreemptorModuleTest.java 67b6d69e3ddd1028dfe9ff451b171cd888674920


Diff: https://reviews.apache.org/r/55357/diff/


Testing
-------

As is, the cluster setup in our existing preemption benchmark does not reflect the improvements
resulting from this patch. Currently, all existing victims can be preempted, therefore all
`PendingTaskProcessor` has to is look at the next slave.

```
BEFORE
Benchmark                                                       (numPendingTasks)   Mode 
Cnt   Score   Error  Units
SchedulingBenchmarks.PreemptorSlotSearchBenchmark.runBenchmark                  1  thrpt 
 10  75.386 ± 2.984  ops/s
SchedulingBenchmarks.PreemptorSlotSearchBenchmark.runBenchmark                 10  thrpt 
 10  74.584 ± 2.598  ops/s
SchedulingBenchmarks.PreemptorSlotSearchBenchmark.runBenchmark                100  thrpt 
 10  79.731 ± 2.182  ops/s
SchedulingBenchmarks.PreemptorSlotSearchBenchmark.runBenchmark               1000  thrpt 
 10  66.386 ± 1.833  ops/s

AFTER
Benchmark                                                       (numPendingTasks)   Mode 
Cnt   Score   Error  Units
SchedulingBenchmarks.PreemptorSlotSearchBenchmark.runBenchmark                  1  thrpt 
 10  78.266 ± 3.290  ops/s
SchedulingBenchmarks.PreemptorSlotSearchBenchmark.runBenchmark                 10  thrpt 
 10  76.743 ± 2.073  ops/s
SchedulingBenchmarks.PreemptorSlotSearchBenchmark.runBenchmark                100  thrpt 
 10  75.343 ± 1.943  ops/s
SchedulingBenchmarks.PreemptorSlotSearchBenchmark.runBenchmark               1000  thrpt 
 10  68.284 ± 2.413  ops/s
```

I need to further imprpve the cluster setup for this benchmark to reflect the improvements
in the patch. A more representative cluster setup would be one in which only a subset of potential
victims pass `PreemptionVictimFilter.filterPreemptionVictims()` test.


Thanks,

Mehrdad Nurolahzade


Mime
  • Unnamed multipart/alternative (inline, None, 0 bytes)
View raw message