Sinece the hadoop task monitor will check each task, and when find it consume 
to much memory, then it will kill the task, so I am currently want to find a 
method to decrease the mem usage at solr side, any idea?
At 2012-01-09 17:07:09,"Tomas Zerolo" <tomas.zer...@axelspringer.de> wrote:
>On Mon, Jan 09, 2012 at 01:29:39PM +0800, James wrote:
>> I am build the solr index on the hadoop, and at reduce step I run the task 
>> that merge the indexes, each part of index is about 1G, I have 10 indexes to 
>> merge them together, I always get the java heap memory exhausted, the heap 
>> size is about 2G  also. I wonder which part use these so many memory. And 
>> how to avoid the OOM during the merge process.
>
>There are three issues in there. You should first try to find out which
>one it is (it's not clear to me based on your question):
>
>  - Java heap memory: you can set that as a start option of the JVM.
>    You set the maximum with the -Xmxn start option. You get an
>    OutOfMemory exception if you reach that (no idea wheter the
>    SOLR code bubbles this up, but there are experts on that here).
>  - Operating system limit: you can set the limit for a process's
>    use of resources (memory, among others). Typically, Linux based
>    systems are shipped with unlimited memory setting; Ralf already
>    posted how to check/set that.
>    The situation here is a bit complicated, because there are
>    different limits (memory size vs. virtual memory size, mainly)
>    and they are exercised differently depending on the allocation
>    pattern. Anyway, I'd expect malloc() returning NULL in this
>    case and the Java runtime translating it (again) into an OutOfMemory
>    exception.
>  - Now the OOM killer is quite another kettle of fish. AFAIK, it's
>    Linux-specific. Once the global system memory is more-or-less
>    exhausted, the kernel kills some applications to try to improve
>    the situation. There's some heuristic in deciding which application
>    to kill, and there are some knobs to help the kernel in this
>    decision. I'd recommend [1]; after reading *that* you know all :-)
>    You know you've run into that by looking at the system log.
>
>
>[1] <https://lwn.net/Articles/317814/>
>-- 
>Tomás Zerolo
>Axel Springer AG
>Axel Springer media Systems
>BILD Produktionssysteme
>Axel-Springer-Straße 65
>10888 Berlin
>Tel.: +49 (30) 2591-72875
>tomas.zer...@axelspringer.de
>www.axelspringer.de
>
>Axel Springer AG, Sitz Berlin, Amtsgericht Charlottenburg, HRB 4998
>Vorsitzender des Aufsichtsrats: Dr. Giuseppe Vita
>Vorstand: Dr. Mathias Döpfner (Vorsitzender)
>Jan Bayer, Ralph Büchi, Lothar Lanz, Dr. Andreas Wiele

Reply via email to