Hi Hao.
Yes, you should be able to instrument any part of Hadoop including
mapreduce daemons. A good examples of how to inject faults to Hadoop
are fault inject tests you can find in Hdfs (under
src/test/aop/org/apache/hadoop). I believe Mapreduce doesn't have any
fault injections tests yet for most of our focus until now was on HDFS
side.
Please keep in mind that you need to write your own aspectj
implementation of the faults (for the weaving) and the you have to
make sure that the instrumented jar files are deployed to the cluster
(e.g. the weaved faults are present in the actual cluster jar files)
Out of the three on second one is in aop.xml.
% ant jar-fault-inject
% ant jar-test-fault-inject
% ant run-test-hdfs-fault-inject
The rest is located in the main build.xml. Pretty ugly I think but I
had to work around certain ant limitation and that was less harmful
workaround of all.
Please feel free to ask further questions if something won't be clear.
--
Take care,
Konstantin (Cos) Boudnik
On Mon, Feb 21, 2011 at 05:30, Hao Zhu peter.zhu...@gmail.com wrote:
Dear cos:
Nice to meet you.
I have a couple of questions about your great job: fault injection
project.
First, is this framework possible to inject a Node-failure fault into
mapreduce? Because i would like to simulate node failure described in
Google's paper in my cluster.
if it possible, please give me some clues to achieve that goal.
Second, i am a New to AOP. So, I just do not quite understand the
following command written in your guide:
% ant jar-fault-inject
% ant jar-test-fault-inject
% ant run-test-hdfs-fault-inject
but i check out the file: hdfs/src/test/aop/build/aop.xml, i
could not find out the corresponding code in that file.
I work a project to run gridmix workload on my cluster with some node's
random failure. So the first question is really important to me.
Finally, look forward to hear from you.
Best Regards
Hao Zhu
