public static class AbstractMRMWTestCase.GenerateExperiment
extends com.bigdata.testutil.ExperimentDriver
The follow result summary is for 20 clients with 80% read tasks (9/27/07). The timeout was 30 and the store files grew up to ~160M (for transient).
bufferMode readBytes/s writeBytes/s read write Transient 22,372,313 5,666,728 Direct 18,560,143 4,823,312 83% 85% Disk 11,315,786 3,001,315 51% 53%
The big puzzle here is why Disk is so much slower than Direct. You can
see that Direct runs at 83% (reads) or 85% (writes) of Transient while
Disk runs at a significantly slow percentage of the Transient
performance. This is odd since both methods are writing on the disk! If
reads are dragging down performance, then maybe read cache for the
DiskOnlyStrategy
would be useful? Note that this class uses
random selection of addresses to read back so this will defeat any cache
once the #of bytes/records written grows large enough.
com.bigdata.testutil.ExperimentDriver.Condition, com.bigdata.testutil.ExperimentDriver.DTDParserHelper, com.bigdata.testutil.ExperimentDriver.Experiment, com.bigdata.testutil.ExperimentDriver.IComparisonTest, com.bigdata.testutil.ExperimentDriver.Result
Constructor and Description |
---|
AbstractMRMWTestCase.GenerateExperiment() |
Modifier and Type | Method and Description |
---|---|
static void |
main(String[] args)
Generates an XML file that can be run by
ExperimentDriver . |
Copyright © 2006–2019 SYSTAP, LLC DBA Blazegraph. All rights reserved.