I haven't done those benchmarks, but if I get some time I will.
The way we run the processes at Visvo, and I will open source the
framework soon, is we have python scripts which run a continuous job
stream of indexing and moving shards of x million pages to search
servers. Each search server has the linkdb, segments, and indexes for
only that shard. We use master crawldb and linkdb but have processes
which will create shard linkdbs for only the content in the shard
segments.
Each shard is pushed out to its own search server. The indexes are
about 2G in size and the segments about 20G. So the tempfs was a quick
hack that turned out to work very well. We could push the shard out and
then run a simple script to mount the indexes in memory. We could also
unmount them back to disk if needed. All of this is transparent to the
SearchServer which thinks it is looking at a local file system.
I do think it would be useful for the SearchServer to have an option for
a RAMDirectory. I don't know if it currently does. This goes into
discussions of creating a master/slave type framework for monitoring and
maintaining shards though.
Dennis Kubes
Enis Soztutar wrote:
Dennis,
Have you tried using o.a.lucene.store.RAMDirectory instead of tempfs.
Intuitively I believe RAMDirectory should be faster, isn't it ? Do you
have any benchmark for the two?
Dennis Kubes wrote:
Trey Spiva wrote:
According to a hadoop tutorial
(http://wiki.apache.org/nutch/NutchHadoopTutorial) on wiki,
"you don't want to search using DFS, you want to search using local
filesystems. Once the index has been created on the DFS you can
use the hadoop copyToLocal command to move it to the local file
system as such" ... "Understand that at this point we are not using
the DFS or MapReduce to do the searching, all of it is on a local
machine".
So my understanding is that hadoop is only good for batch index
building, and is not proper for incremental index building and
search. Is this true?
That is correct. DFS for batch processing and MapReduce jobs. Local
servers (disks) for serving indexes. Even better put local indexes
(not segments, just indexes) in RAM.
The reason I am asking is that when I read the article ACM article by
Mike Cafarella and Doug Cutting, to me it sounded like the concern
was to make the index structures fit in the primary memory, not the
entire crawled database. Did I miss understand the ACM article?
No, what they are saying is the more pages per index per machine on
hard disk the slower the search. Keeping the main indexes, but not
the segments which hold raw page content, in RAM can speed up search
significantly.
One way to do this if you are running on linux is to create a tempfs
(which is ram) and then mount the filesystem in the ram. Then your
index acts normally to the application but is essentially served from
Ram. This is how we server the Nutch lucene indexes on our web search
engine (www.visvo.com) which is ~100M pages. Below is how you can
achieve this, assuming your indexes are in /path/to/indexes:
mv /path/to/indexes /path/to/indexes.dist
mkdir /path/to/indexes
cd /path/to
mount -t tmpfs -o size=2684354560 none /path/to/indexes
rsync --progress -aptv indexes.dist/* indexes/
chown -R user:group indexes
This would of course be limited by the amount of RAM you have on the
machine. But with this approach most searches are sub-second.
Dennis Kubes
